[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"articles":3,"summaries-facets-categories":5260},[4,626,1799,2412,3224,3389,3542,3935,4145],{"id":5,"title":6,"ai_summary":7,"author":8,"body":9,"category":600,"date_modified":601,"description":602,"excerpt":8,"extension":603,"faq":601,"keywords":604,"meta":613,"meta_description":614,"meta_title":615,"navigation":616,"path":617,"primary_keyword":605,"published_at":601,"related_posts":618,"sections":619,"seo":620,"sitemap":621,"slug":622,"status":623,"stem":624,"__hash__":625},"articles\u002Farticles\u002Fagent-architecture-the-orchestration-stack-that-actually-emerged-article.md","The Emergent AI Agent Orchestration Stack: Harnesses, Specs, and Primitives","AI agent orchestration stacks mature compute layers atop lagging coordination using YAML harnesses, spec-driven workflows, and primitives like tool registries and state persistence. Tools such as Archon V3 and Claude Code enable deterministic multi-agent execution for production reliability without frameworks.","",{"type":10,"value":11,"toc":591},"minimark",[12,35,45,50,59,69,81,84,94,98,109,119,129,132,136,147,163,166,169,173,182,204,221,224,228,237,257,260,263,323,326,329,333,344,361,384,387,390,587],[13,14,15,16,27],"p",{},"Production AI agent orchestration demands stack literacy across six uneven layers, from mature compute sandboxes like E2B Firecracker to lagging coordination without Kubernetes-grade tools. Builders bridge gaps with YAML harnesses and spec-driven workflows today, delivering deterministic multi-agent execution via Archon V3 and Claude Code primitives.",[17,18,19],"sup",{},[20,21,26],"a",{"href":22,"ariaDescribedBy":23,"dataFootnoteRef":8,"id":25},"#user-content-fn-1",[24],"footnote-label","user-content-fnref-1","1",[17,28,29],{},[20,30,34],{"href":31,"ariaDescribedBy":32,"dataFootnoteRef":8,"id":33},"#user-content-fn-2",[24],"user-content-fnref-2","2",[13,36,37,38,44],{},"Demos shine with single agents, but real systems crumble on orchestration failures—compounding errors across layers drop end-to-end reliability from five 99% components to 95% overall.",[17,39,40],{},[20,41,26],{"href":22,"ariaDescribedBy":42,"dataFootnoteRef":8,"id":43},[24],"user-content-fnref-1-2"," Without primitives like tool registries and state persistence, you risk hyperscaler lock-in and sprawl. This guide maps the stack, harnesses, specs, challenges, and pitfalls to ship reliable agents now.",[46,47,49],"h2",{"id":48},"the-layered-maturity-of-the-ai-agent-stack","The Layered Maturity of the AI Agent Stack",[13,51,52,53],{},"The AI agent stack matures unevenly across six layers—compute and sandboxing lead with mature tools like Browserbase and E2B Firecracker for isolated execution, while orchestration lags like pre-Kubernetes eras without infra-grade scheduling or FinOps. Identity\u002Fcomms transitions via emerging agent-native protocols beyond email shims. This creates production bottlenecks in coordination, demanding immediate stack literacy.",[17,54,55],{},[20,56,26],{"href":22,"ariaDescribedBy":57,"dataFootnoteRef":8,"id":58},[24],"user-content-fnref-1-3",[13,60,61,62,68],{},"Nate B. Jones breaks it into six layers in his analysis.",[17,63,64],{},[20,65,26],{"href":22,"ariaDescribedBy":66,"dataFootnoteRef":8,"id":67},[24],"user-content-fnref-1-4"," Compute\u002Fsandboxing handles isolated execution reliably. Identity\u002Fcomms transitions with emerging agent-native protocols over email shims.",[13,70,71,72,80],{},"Memory sits early, blending Mem0 hybrids but risking vendor lock-in.",[17,73,74],{},[20,75,79],{"href":76,"ariaDescribedBy":77,"dataFootnoteRef":8,"id":78},"#user-content-fn-3",[24],"user-content-fnref-3","3"," Tools explode via Compose connectors for auth-heavy integrations.",[13,82,83],{},"Provisioning emerges in Stripe Projects for dynamic scaling. Orchestration gaps lack FinOps and infra-grade controls—no standard for fleets.",[13,85,86,87,93],{},"Ephemeral agents suit quick tasks; persistent ones need state for long runs. Gartner notes a 1,445% multi-agent surge, amplifying these mismatches.",[17,88,89],{},[20,90,26],{"href":22,"ariaDescribedBy":91,"dataFootnoteRef":8,"id":92},[24],"user-content-fnref-1-5"," Skip stack literacy, and your agents stay demo-bound.",[46,95,97],{"id":96},"what-is-harness-engineering-for-ai-agent-orchestration","What Is Harness Engineering for AI Agent Orchestration?",[13,99,100,101],{},"Harness engineering wraps unreliable AI agents in declarative YAML or Markdown workflows that use DAGs for dependencies, Git worktrees for parallel isolation, and pre\u002Fpost-tool hooks for self-correction and verification loops like type-checks or rewrites. This enables Stripe-scale PRs (1,300\u002Fweek) without constant oversight by treating agents as deterministic nodes.",[17,102,103],{},[20,104,108],{"href":105,"ariaDescribedBy":106,"dataFootnoteRef":8,"id":107},"#user-content-fn-4",[24],"user-content-fnref-4","4",[13,110,111,112,118],{},"Archon V3 defines this: YAML commands like classify\u002Fplan\u002Fimplement form nodes in DAGs with dependencies.",[17,113,114],{},[20,115,108],{"href":105,"ariaDescribedBy":116,"dataFootnoteRef":8,"id":117},[24],"user-content-fnref-4-2"," Worktrees isolate parallel runs—four agents at once without clashes. Pre\u002Fpost-tool hooks loop for verification, like type-checks or rewrites.",[13,120,121,122,128],{},"Claude Code adds structured metadata, permissions, and multi-agent forks.",[17,123,124],{},[20,125,34],{"href":31,"ariaDescribedBy":126,"dataFootnoteRef":8,"id":127},[24],"user-content-fnref-2-2"," Mix precise steps with AI nodes; extend as Markdown primitives standard in Anthropic\u002FOpenAI for chaining. You treat agents as cogs in deterministic machines, not free-range thinkers.",[13,130,131],{},"This beats prompt tweaks alone. Version YAML in Git for audits. Production demands it over ad-hoc calls.",[46,133,135],{"id":134},"the-shift-to-spec-centric-development","The Shift to Spec-Centric Development",[13,137,138,139],{},"Spec-centric development flips code-first workflows by making declarative specs in YAML or Markdown the executable single source of truth, driving AI agents like GitHub SpecKit to generate synchronized code through staged pipelines of specify\u002Fplan\u002Ftasks\u002Fimplement for consistent handoffs.",[17,140,141],{},[20,142,146],{"href":143,"ariaDescribedBy":144,"dataFootnoteRef":8,"id":145},"#user-content-fn-5",[24],"user-content-fnref-5","5",[13,148,149,150,156,157],{},"Code drifts specs outdated; specs drive regeneration instead.",[17,151,152],{},[20,153,146],{"href":143,"ariaDescribedBy":154,"dataFootnoteRef":8,"id":155},[24],"user-content-fnref-5-2"," GitHub SpecKit structures .github\u002Fprompts\u002Fagents for PM\u002Farchitect\u002Fengineer handoffs. Anthropic harnesses use minimal three agents—planner\u002Fgenerator\u002Fevaluator—outpacing bloated frameworks on Opus models.",[17,158,159],{},[20,160,34],{"href":31,"ariaDescribedBy":161,"dataFootnoteRef":8,"id":162},[24],"user-content-fnref-2-3",[13,164,165],{},"Version specs in repos for team reliability. No more merge hell from AI code dumps. Specs ensure handoffs work; code follows.",[13,167,168],{},"I prefer this for complex projects. It cuts ambiguity. Agents execute living docs, not stale prose.",[46,170,172],{"id":171},"why-ai-agent-orchestration-is-the-biggest-unsolved-problem","Why AI Agent Orchestration Is the Biggest Unsolved Problem",[13,174,175,176],{},"AI agent orchestration fails production due to critical gaps in scheduling\u002Flifecycle management, supervision hierarchies, FinOps controls, inter-agent comms protocols, and observability tools—current libraries like LangGraph manage notebooks but crumble under 50-agent fleets needing audits and traces.",[17,177,178],{},[20,179,79],{"href":76,"ariaDescribedBy":180,"dataFootnoteRef":8,"id":181},[24],"user-content-fnref-3-2",[183,184,185,195,198,201],"ul",{},[186,187,188,189],"li",{},"Dynamic scheduling beats cron; state persists across crashes.",[17,190,191],{},[20,192,79],{"href":76,"ariaDescribedBy":193,"dataFootnoteRef":8,"id":194},[24],"user-content-fnref-3-3",[186,196,197],{},"Retry logic and error isolation prevent cascades.",[186,199,200],{},"Token budgeting enforces FinOps per task.",[186,202,203],{},"MCP protocols limit comms; no standard for negotiation.",[13,205,206,207,213,214,220],{},"MindStudio calls it the core blocker: demos route simply, production demands hierarchies and traces.",[17,208,209],{},[20,210,79],{"href":76,"ariaDescribedBy":211,"dataFootnoteRef":8,"id":212},[24],"user-content-fnref-3-4"," Jones likens it to pre-K8s chaos.",[17,215,216],{},[20,217,26],{"href":22,"ariaDescribedBy":218,"dataFootnoteRef":8,"id":219},[24],"user-content-fnref-1-6"," Flexibility invites emergence; determinism suits enterprise.",[13,222,223],{},"Trade-off bites: emergent behavior thrills but flakes. Prioritize primitives over frameworks. Libraries add overhead without infra.",[46,225,227],{"id":226},"key-primitives-for-production-ai-agent-orchestration","Key Primitives for Production AI Agent Orchestration",[13,229,230,231],{},"Production AI agent orchestration relies on 12 primitives from Claude Code leaks, including tool registries for metadata filtering, tiered permissions to block rogue actions, state persistence via JSON sessions, token budgeting with projections and halts, structured logging through typed events, verification loops, plus agent typing and dynamic pools for crash-resilient, observable multi-agent workflows.",[17,232,233],{},[20,234,34],{"href":31,"ariaDescribedBy":235,"dataFootnoteRef":8,"id":236},[24],"user-content-fnref-2-4",[183,238,239,248,251,254],{},[186,240,241,242],{},"Dual registries list 207 commands with metadata for filtering.",[17,243,244],{},[20,245,34],{"href":31,"ariaDescribedBy":246,"dataFootnoteRef":8,"id":247},[24],"user-content-fnref-2-5",[186,249,250],{},"Permissions tier built-in\u002Fplugins\u002Fuser to block rogue acts.",[186,252,253],{},"JSON persists sessions; projections halt overruns.",[186,255,256],{},"Typed events stream logs.",[13,258,259],{},"Agents specialize: explore\u002Fplan\u002Fverify types in hierarchies. Formula: Role+Goal+Tools+Rules+Output.",[13,261,262],{},"Code stub for tools:",[264,265,269],"pre",{"className":266,"code":267,"language":268,"meta":8,"style":8},"language-yaml shiki shiki-themes github-light github-dark","tools:\n  - name: list_files\n    description: List files in directory\n    source: builtin\n","yaml",[270,271,272,285,301,312],"code",{"__ignoreMap":8},[273,274,277,281],"span",{"class":275,"line":276},"line",1,[273,278,280],{"class":279},"s9eBZ","tools",[273,282,284],{"class":283},"sVt8B",":\n",[273,286,288,291,294,297],{"class":275,"line":287},2,[273,289,290],{"class":283},"  - ",[273,292,293],{"class":279},"name",[273,295,296],{"class":283},": ",[273,298,300],{"class":299},"sZZnC","list_files\n",[273,302,304,307,309],{"class":275,"line":303},3,[273,305,306],{"class":279},"    description",[273,308,296],{"class":283},[273,310,311],{"class":299},"List files in directory\n",[273,313,315,318,320],{"class":275,"line":314},4,[273,316,317],{"class":279},"    source",[273,319,296],{"class":283},[273,321,322],{"class":299},"builtin\n",[13,324,325],{},"Persist post-events. Update claude.md for skills; \u002Fplan for tasks. Pools assemble per-session.",[13,327,328],{},"These make fleets observable. Skip them, stay small-scale.",[46,330,332],{"id":331},"common-pitfalls-and-trade-offs-in-multi-agent-systems","Common Pitfalls and Trade-offs in Multi-Agent Systems",[13,334,335,336],{},"Multi-agent systems trip on overkill by jumping to agents before mastering 5 workflows, bloat from skills exceeding 150 lines, lock-in via hyperscaler memory, and ignoring compounding failures that drop reliability—counter them by testing quantitatively, tiering skills, and designing agent-first for composability.",[17,337,338],{},[20,339,343],{"href":340,"ariaDescribedBy":341,"dataFootnoteRef":8,"id":342},"#user-content-fn-6",[24],"user-content-fnref-6","6",[183,345,346,355,358],{},[186,347,348,349],{},"Vague skills undertrigger; self-evals overconfident; no permissions equals demos.",[17,350,351],{},[20,352,34],{"href":31,"ariaDescribedBy":353,"dataFootnoteRef":8,"id":354},[24],"user-content-fnref-2-6",[186,356,357],{},"Single agents scale until they don't—multi for handoffs.",[186,359,360],{},"Ephemeral for bursts; persistent for state.",[13,362,363,364,370,371,377,378],{},"Frameworks bloat post-Opus; chaining routes first.",[17,365,366],{},[20,367,343],{"href":340,"ariaDescribedBy":368,"dataFootnoteRef":8,"id":369},[24],"user-content-fnref-6-2"," Margerie: Master Role+Goal+Tools+Rules+Output.",[17,372,373],{},[20,374,343],{"href":340,"ariaDescribedBy":375,"dataFootnoteRef":8,"id":376},[24],"user-content-fnref-6-3"," Jones: Primitives avert pain.",[17,379,380],{},[20,381,26],{"href":22,"ariaDescribedBy":382,"dataFootnoteRef":8,"id":383},[24],"user-content-fnref-1-7",[13,385,386],{},"Test E2E: 99% layers compound poorly. Agent-first composes; code-first breaks.",[13,388,389],{},"Audit your agent workflow against Claude Code's 12 primitives today. Fork an Archon V3 YAML harness into a Git repo, route a test issue through planner\u002Fgenerator\u002Fevaluator, and log E2E reliability metrics before adding agents.",[391,392,395,400],"section",{"className":393,"dataFootnotes":8},[394],"footnotes",[46,396,399],{"className":397,"id":24},[398],"sr-only","Footnotes",[401,402,403,458,502,532,548,564],"ol",{},[186,404,406,407,414,415,414,422,414,429,414,436,414,443,414,450],{"id":405},"user-content-fn-1","Nate B. Jones, \"6-Layer AI Agent Stack: Build Literacy Now,\" AI News & Strategy Daily. ",[20,408,413],{"href":409,"ariaLabel":410,"className":411,"dataFootnoteBackref":8},"#user-content-fnref-1","Back to reference 1",[412],"data-footnote-backref","↩"," ",[20,416,413,420],{"href":417,"ariaLabel":418,"className":419,"dataFootnoteBackref":8},"#user-content-fnref-1-2","Back to reference 1-2",[412],[17,421,34],{},[20,423,413,427],{"href":424,"ariaLabel":425,"className":426,"dataFootnoteBackref":8},"#user-content-fnref-1-3","Back to reference 1-3",[412],[17,428,79],{},[20,430,413,434],{"href":431,"ariaLabel":432,"className":433,"dataFootnoteBackref":8},"#user-content-fnref-1-4","Back to reference 1-4",[412],[17,435,108],{},[20,437,413,441],{"href":438,"ariaLabel":439,"className":440,"dataFootnoteBackref":8},"#user-content-fnref-1-5","Back to reference 1-5",[412],[17,442,146],{},[20,444,413,448],{"href":445,"ariaLabel":446,"className":447,"dataFootnoteBackref":8},"#user-content-fnref-1-6","Back to reference 1-6",[412],[17,449,343],{},[20,451,413,455],{"href":452,"ariaLabel":453,"className":454,"dataFootnoteBackref":8},"#user-content-fnref-1-7","Back to reference 1-7",[412],[17,456,457],{},"7",[186,459,461,462,414,467,414,474,414,481,414,488,414,495],{"id":460},"user-content-fn-2","Nick Puru, \"Claude Code Leak Reveals Full AI Orchestration Engine,\" AI Automation. ",[20,463,413],{"href":464,"ariaLabel":465,"className":466,"dataFootnoteBackref":8},"#user-content-fnref-2","Back to reference 2",[412],[20,468,413,472],{"href":469,"ariaLabel":470,"className":471,"dataFootnoteBackref":8},"#user-content-fnref-2-2","Back to reference 2-2",[412],[17,473,34],{},[20,475,413,479],{"href":476,"ariaLabel":477,"className":478,"dataFootnoteBackref":8},"#user-content-fnref-2-3","Back to reference 2-3",[412],[17,480,79],{},[20,482,413,486],{"href":483,"ariaLabel":484,"className":485,"dataFootnoteBackref":8},"#user-content-fnref-2-4","Back to reference 2-4",[412],[17,487,108],{},[20,489,413,493],{"href":490,"ariaLabel":491,"className":492,"dataFootnoteBackref":8},"#user-content-fnref-2-5","Back to reference 2-5",[412],[17,494,146],{},[20,496,413,500],{"href":497,"ariaLabel":498,"className":499,"dataFootnoteBackref":8},"#user-content-fnref-2-6","Back to reference 2-6",[412],[17,501,343],{},[186,503,505,506,414,511,414,518,414,525],{"id":504},"user-content-fn-3","MindStudio, \"What Is Agent Orchestration? Why It's the Biggest Unsolved Problem in the AI Stack.\" ",[20,507,413],{"href":508,"ariaLabel":509,"className":510,"dataFootnoteBackref":8},"#user-content-fnref-3","Back to reference 3",[412],[20,512,413,516],{"href":513,"ariaLabel":514,"className":515,"dataFootnoteBackref":8},"#user-content-fnref-3-2","Back to reference 3-2",[412],[17,517,34],{},[20,519,413,523],{"href":520,"ariaLabel":521,"className":522,"dataFootnoteBackref":8},"#user-content-fnref-3-3","Back to reference 3-3",[412],[17,524,79],{},[20,526,413,530],{"href":527,"ariaLabel":528,"className":529,"dataFootnoteBackref":8},"#user-content-fnref-3-4","Back to reference 3-4",[412],[17,531,108],{},[186,533,535,536,414,541],{"id":534},"user-content-fn-4","\"Archon V3: YAML Harnesses for AI Coding Agents,\" DIY Smart Code. ",[20,537,413],{"href":538,"ariaLabel":539,"className":540,"dataFootnoteBackref":8},"#user-content-fnref-4","Back to reference 4",[412],[20,542,413,546],{"href":543,"ariaLabel":544,"className":545,"dataFootnoteBackref":8},"#user-content-fnref-4-2","Back to reference 4-2",[412],[17,547,34],{},[186,549,551,552,414,557],{"id":550},"user-content-fn-5","\"SDD Makes Specs the Single Source of Truth via AI Agents,\" Level Up Coding. ",[20,553,413],{"href":554,"ariaLabel":555,"className":556,"dataFootnoteBackref":8},"#user-content-fnref-5","Back to reference 5",[412],[20,558,413,562],{"href":559,"ariaLabel":560,"className":561,"dataFootnoteBackref":8},"#user-content-fnref-5-2","Back to reference 5-2",[412],[17,563,34],{},[186,565,567,568,414,573,414,580],{"id":566},"user-content-fn-6","Lukas Margerie insights on agent fundamentals. ",[20,569,413],{"href":570,"ariaLabel":571,"className":572,"dataFootnoteBackref":8},"#user-content-fnref-6","Back to reference 6",[412],[20,574,413,578],{"href":575,"ariaLabel":576,"className":577,"dataFootnoteBackref":8},"#user-content-fnref-6-2","Back to reference 6-2",[412],[17,579,34],{},[20,581,413,585],{"href":582,"ariaLabel":583,"className":584,"dataFootnoteBackref":8},"#user-content-fnref-6-3","Back to reference 6-3",[412],[17,586,79],{},[588,589,590],"style",{},"html pre.shiki code .s9eBZ, html code.shiki .s9eBZ{--shiki-default:#22863A;--shiki-dark:#85E89D}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":8,"searchDepth":287,"depth":287,"links":592},[593,594,595,596,597,598,599],{"id":48,"depth":287,"text":49},{"id":96,"depth":287,"text":97},{"id":134,"depth":287,"text":135},{"id":171,"depth":287,"text":172},{"id":226,"depth":287,"text":227},{"id":331,"depth":287,"text":332},{"id":24,"depth":287,"text":399},"ai-agents",null,"Production AI agent orchestration demands stack literacy across six uneven layers, from mature compute sandboxes like E2B Firecracker to lagging coordination without Kubernetes-grade tools. Builders bridge gaps with YAML harnesses and spec-driven workflows today, delivering deterministic multi-agent execution via Archon V3 and Claude Code primitives.12","md",[605,606,607,608,609,610,611,612],"AI agent orchestration","agent stack","harness engineering","spec-centric development","multi-agent systems","Claude Code","Archon V3","production AI agents",{},"Master the emergent AI agent orchestration stack: uneven maturity, harness engineering, spec-centric dev, and key primitives for production multi-agent systems. Avoid pitfalls with practical workflows from Claude Code and Archon V3.","AI Agent Orchestration Stack: Layers and Primitives",true,"\u002Farticles\u002Fagent-architecture-the-orchestration-stack-that-actually-emerged-article",[],{"the-layered-maturity-of-the-ai-agent-stack":49,"what-is-harness-engineering-for-ai-agent-orchestration":97,"the-shift-to-spec-centric-development":135,"why-ai-agent-orchestration-is-the-biggest-unsolved-problem":172,"key-primitives-for-production-ai-agent-orchestration":227,"common-pitfalls-and-trade-offs-in-multi-agent-systems":332},{"title":6,"description":602},{"loc":617},"emergent-ai-agent-orchestration-stack-harnesess-specs-primitives","ready","articles\u002Fagent-architecture-the-orchestration-stack-that-actually-emerged-article","xdOSx_FUsF7EEM6WqQRjgq4p4T-5gHkBnwndjo2dz-M",{"id":627,"title":628,"ai_summary":629,"author":8,"body":630,"category":1779,"date_modified":601,"description":1780,"excerpt":8,"extension":603,"faq":601,"keywords":1781,"meta":1787,"meta_description":1788,"meta_title":1789,"navigation":616,"path":1790,"primary_keyword":638,"published_at":601,"related_posts":1791,"sections":1792,"seo":1794,"sitemap":1795,"slug":1796,"status":623,"stem":1797,"__hash__":1798},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-article.md","The 3-Core Agent Harness: Planner, Generator, Evaluator for Production AI Agents","Production AI agent systems succeed with a 3-core agent harness: Planner creates high-level outlines, Generator builds outputs from them, and Evaluator critiques via graded rubrics. This strips 90% redundant framework overhead, leveraging modern LLMs for reliable, scalable results on long tasks.",{"type":10,"value":631,"toc":1769},[632,646,649,658,666,680,683,891,894,898,907,910,918,932,935,939,950,957,965,973,981,994,997,1006,1009,1045,1048,1052,1062,1065,1068,1071,1080,1083,1087,1099,1102,1105,1113,1116,1120,1133,1163,1171,1177,1180,1255,1258,1264,1291,1294,1303,1309,1389,1404,1409,1441,1453,1456,1460,1476,1479,1503,1506,1515,1518,1521,1525,1534,1537,1563,1573,1576,1579,1582,1766],[13,633,634,635,639,640,645],{},"Production AI agent systems thrive with a ",[636,637,638],"strong",{},"3-core agent harness","—Planner for high-level outlines, Generator for implementation, Evaluator for adversarial critique using graded rubrics—stripping 90% bloated framework overhead per Anthropic tests.",[17,641,642],{},[20,643,26],{"href":22,"ariaDescribedBy":644,"dataFootnoteRef":8,"id":25},[24]," This leverages modern LLMs' 1M+ token windows and coherence for reliable, scalable outputs on long-horizon tasks.",[13,647,648],{},"You've likely tried agent frameworks that promise reliability but deliver cascading errors from micro-task sharding. Those assumptions fit older LLMs with short contexts and poor coherence, not Claude Opus 4.6. This harness ships product-level work like high-volume AI-generated PRs, letting you focus on primitives over complexity.",[13,650,651,652,657],{},"Anthropic's experiments, summarized by AI LABS, tested stripping frameworks layer by layer on long tasks.",[17,653,654],{},[20,655,26],{"href":22,"ariaDescribedBy":656,"dataFootnoteRef":8,"id":43},[24]," Results showed only the Planner for product outlines, Generator for execution, and Evaluator for critique delivered gains. Complex staging in SpecKit or loops in GSD added token bloat and error drift without benefits.",[13,659,660,661],{},"Contrast traditional setups: BMAD shards PRDs into micro-tasks to handle context resets; GSD enforces rigid agent loops for weak reasoning. With 1M-token windows, LLMs now manage full product scopes autonomously—micro-sharding forces premature decisions that fail downstream.",[17,662,663],{},[20,664,26],{"href":22,"ariaDescribedBy":665,"dataFootnoteRef":8,"id":58},[24],[13,667,668,669,674,675],{},"Nate B. Jones's analysis of Claude's leaked code reinforces primitives over frameworks for scalability.",[17,670,671],{},[20,672,34],{"href":31,"ariaDescribedBy":673,"dataFootnoteRef":8,"id":33},[24]," Agent Blueprint advises single-agent mastery before multi-agent scaling.",[17,676,677],{},[20,678,79],{"href":76,"ariaDescribedBy":679,"dataFootnoteRef":8,"id":78},[24],[13,681,682],{},"To see it in action, here's a minimal YAML harness stub using Archon V3 style for a task-tracking app:",[264,684,686],{"className":266,"code":685,"language":268,"meta":8,"style":8},"harness:\n  planner:\n    role: \"Product Lead\"\n    goal: \"Outline task-tracking app\"\n    tools: [\"search\"]\n    rules: [\"High-level only; flag risks; no code\"]\n    output: \"Markdown: user stories, features, roadmap\"\n  generator:\n    role: \"Engineer\"\n    goal: \"Implement planner outline\"\n    tools: [\"git\", \"calc\"]\n    rules: [\"Isolated worktree; commit artifacts\"]\n    output: \"Code files + README\"\n  evaluator:\n    role: \"QA Adversary\"\n    goal: \"Score vs rubric\"\n    tools: [\"playwright\"]\n    rules: [\"Assume flaws; weighted scores\"]\n    output: \"JSON: scores, fixes\"\n",[270,687,688,695,702,712,722,737,750,761,769,779,789,807,819,829,837,847,857,869,881],{"__ignoreMap":8},[273,689,690,693],{"class":275,"line":276},[273,691,692],{"class":279},"harness",[273,694,284],{"class":283},[273,696,697,700],{"class":275,"line":287},[273,698,699],{"class":279},"  planner",[273,701,284],{"class":283},[273,703,704,707,709],{"class":275,"line":303},[273,705,706],{"class":279},"    role",[273,708,296],{"class":283},[273,710,711],{"class":299},"\"Product Lead\"\n",[273,713,714,717,719],{"class":275,"line":314},[273,715,716],{"class":279},"    goal",[273,718,296],{"class":283},[273,720,721],{"class":299},"\"Outline task-tracking app\"\n",[273,723,725,728,731,734],{"class":275,"line":724},5,[273,726,727],{"class":279},"    tools",[273,729,730],{"class":283},": [",[273,732,733],{"class":299},"\"search\"",[273,735,736],{"class":283},"]\n",[273,738,740,743,745,748],{"class":275,"line":739},6,[273,741,742],{"class":279},"    rules",[273,744,730],{"class":283},[273,746,747],{"class":299},"\"High-level only; flag risks; no code\"",[273,749,736],{"class":283},[273,751,753,756,758],{"class":275,"line":752},7,[273,754,755],{"class":279},"    output",[273,757,296],{"class":283},[273,759,760],{"class":299},"\"Markdown: user stories, features, roadmap\"\n",[273,762,764,767],{"class":275,"line":763},8,[273,765,766],{"class":279},"  generator",[273,768,284],{"class":283},[273,770,772,774,776],{"class":275,"line":771},9,[273,773,706],{"class":279},[273,775,296],{"class":283},[273,777,778],{"class":299},"\"Engineer\"\n",[273,780,782,784,786],{"class":275,"line":781},10,[273,783,716],{"class":279},[273,785,296],{"class":283},[273,787,788],{"class":299},"\"Implement planner outline\"\n",[273,790,792,794,796,799,802,805],{"class":275,"line":791},11,[273,793,727],{"class":279},[273,795,730],{"class":283},[273,797,798],{"class":299},"\"git\"",[273,800,801],{"class":283},", ",[273,803,804],{"class":299},"\"calc\"",[273,806,736],{"class":283},[273,808,810,812,814,817],{"class":275,"line":809},12,[273,811,742],{"class":279},[273,813,730],{"class":283},[273,815,816],{"class":299},"\"Isolated worktree; commit artifacts\"",[273,818,736],{"class":283},[273,820,822,824,826],{"class":275,"line":821},13,[273,823,755],{"class":279},[273,825,296],{"class":283},[273,827,828],{"class":299},"\"Code files + README\"\n",[273,830,832,835],{"class":275,"line":831},14,[273,833,834],{"class":279},"  evaluator",[273,836,284],{"class":283},[273,838,840,842,844],{"class":275,"line":839},15,[273,841,706],{"class":279},[273,843,296],{"class":283},[273,845,846],{"class":299},"\"QA Adversary\"\n",[273,848,850,852,854],{"class":275,"line":849},16,[273,851,716],{"class":279},[273,853,296],{"class":283},[273,855,856],{"class":299},"\"Score vs rubric\"\n",[273,858,860,862,864,867],{"class":275,"line":859},17,[273,861,727],{"class":279},[273,863,730],{"class":283},[273,865,866],{"class":299},"\"playwright\"",[273,868,736],{"class":283},[273,870,872,874,876,879],{"class":275,"line":871},18,[273,873,742],{"class":279},[273,875,730],{"class":283},[273,877,878],{"class":299},"\"Assume flaws; weighted scores\"",[273,880,736],{"class":283},[273,882,884,886,888],{"class":275,"line":883},19,[273,885,755],{"class":279},[273,887,296],{"class":283},[273,889,890],{"class":299},"\"JSON: scores, fixes\"\n",[13,892,893],{},"Run via GitHub Actions: trigger on issue, parse YAML, invoke Claude sequentially. This beats npm installs—I've shipped PRs weekly this way. We'll detail roles, primitives, build steps, and trade-offs next.",[46,895,897],{"id":896},"why-production-agent-systems-need-a-3-core-agent-harness","Why Production Agent Systems Need a 3-Core Agent Harness",[13,899,900,901,906],{},"Ditch bloated agent frameworks—a 3-core agent harness with Planner, Generator, and Evaluator outperforms them by playing to LLM strengths like large contexts and autonomy. Anthropic's Claude Opus 4.6 tests show 90% of components in BMAD, GSD, and SpecKit are redundant overhead that propagates errors on long-horizon tasks.",[17,902,903],{},[20,904,26],{"href":22,"ariaDescribedBy":905,"dataFootnoteRef":8,"id":67},[24]," High-level planning avoids flawed detailed specs, while separate evaluation enforces rigor.",[13,908,909],{},"Traditional frameworks compensate for outdated LLM limits: context resets demand sharding, weak coherence needs rigid loops. But 1M-token windows and improved reasoning make micro-tasking counterproductive—it forces premature decisions that cascade failures.",[13,911,912,913],{},"Anthropic's experiments, summarized by AI LABS, tested stripping frameworks layer by layer. Results? Only the planner for product outlines, generator for execution, and evaluator for critique yielded gains. Complex staging in SpecKit or GSD loops adds no value; it bloats token use and invites drift.",[17,914,915],{},[20,916,26],{"href":22,"ariaDescribedBy":917,"dataFootnoteRef":8,"id":92},[24],[13,919,920,921,926,927],{},"Nate B. Jones's analysis of Claude's leaked code echoes this: primitives beat frameworks for scalability.",[17,922,923],{},[20,924,34],{"href":31,"ariaDescribedBy":925,"dataFootnoteRef":8,"id":127},[24]," Agent Blueprint stresses single-agent mastery first, scaling only for distinct skills.",[17,928,929],{},[20,930,79],{"href":76,"ariaDescribedBy":931,"dataFootnoteRef":8,"id":181},[24],[13,933,934],{},"I've built agents both ways. Frameworks feel safe until production hits—then you're debugging orchestration bugs, not shipping features. The harness cuts that noise.",[46,936,938],{"id":937},"the-planners-role-in-a-3-core-agent-harness","The Planner's Role in a 3-Core Agent Harness",[13,940,941,942,944,945],{},"In a ",[636,943,638],{},", the Planner generates high-level product outlines, feature breakdowns, and user stories—not micro-tasks—enabling LLMs to autonomously discover optimal paths and avoid error cascades from over-specification. Anthropic's boundary-testing prompts prove Claude Opus 4.6 excels here, generating phased docs without diving into code.",[17,946,947],{},[20,948,26],{"href":22,"ariaDescribedBy":949,"dataFootnoteRef":8,"id":219},[24],[13,951,952,953,956],{},"Start prompts with ",[636,954,955],{},"Role + Goal + high-level deliverables",". Here's a concrete example for a task-tracking app:",[264,958,963],{"className":959,"code":961,"language":962},[960],"language-text","You are a product lead with 10+ years shipping SaaS.\n\nGoal: Outline a task-tracking app for indie devs.\n\nDeliverables:\n- 5-10 user stories (As a [user], I want [feature] so [benefit])\n- Key features prioritized by MVP\n- Phased roadmap (Week 1: Core; Week 2: Polish)\n- Risks and assumptions\n\nRules: Stay high-level—no code, tech stacks, or UI mocks. Flag unknowns.\n\nOutput: Markdown sections only.\n","text",[270,964,961],{"__ignoreMap":8},[13,966,967,968],{},"Claude Opus 4.6 generates phased docs like this, iterating on creative paths humans might miss.",[17,969,970],{},[20,971,26],{"href":22,"ariaDescribedBy":972,"dataFootnoteRef":8,"id":383},[24],[13,974,975,976],{},"Avoid native \"plan\" modes—they fixate on implementation details too early. Archon V3 uses declarative YAML for living workflows, updating plans mid-run.",[17,977,978],{},[20,979,108],{"href":105,"ariaDescribedBy":980,"dataFootnoteRef":8,"id":107},[24],[13,982,983,984,987,988,993],{},"Agent Blueprint's formula drives reliability: ",[636,985,986],{},"Role + Goal + Tools + Rules + Output",".",[17,989,990],{},[20,991,79],{"href":76,"ariaDescribedBy":992,"dataFootnoteRef":8,"id":194},[24]," For Planner: Tools minimal (search only), Rules (\"Stay high-level; flag risks\"), Output (Markdown sections).",[13,995,996],{},"Benefits compound. LLMs outperform humans on detail synthesis—let them handle it post-outline. I've used this for 20+ newsletter outlines; it surfaces non-obvious features like \"AI-suggested subtasks\" I overlooked.",[13,998,999,1000,1005],{},"Contrast SpecKit: It stages detailed specs early, locking in assumptions that drift during generation.",[17,1001,1002],{},[20,1003,146],{"href":143,"ariaDescribedBy":1004,"dataFootnoteRef":8,"id":145},[24]," High-level planning iterates faster.",[13,1007,1008],{},"For code integration, pipe output to Generator via file:",[264,1010,1014],{"className":1011,"code":1012,"language":1013,"meta":8,"style":8},"language-bash shiki shiki-themes github-light github-dark","claude --prompt planner_prompt.md --output planner_outline.md\ngit add planner_outline.md\n","bash",[270,1015,1016,1035],{"__ignoreMap":8},[273,1017,1018,1022,1026,1029,1032],{"class":275,"line":276},[273,1019,1021],{"class":1020},"sScJk","claude",[273,1023,1025],{"class":1024},"sj4cs"," --prompt",[273,1027,1028],{"class":299}," planner_prompt.md",[273,1030,1031],{"class":1024}," --output",[273,1033,1034],{"class":299}," planner_outline.md\n",[273,1036,1037,1040,1043],{"class":275,"line":287},[273,1038,1039],{"class":1020},"git",[273,1041,1042],{"class":299}," add",[273,1044,1034],{"class":299},[13,1046,1047],{},"This keeps plans as artifacts, versioned and reviewable. Scale to repos: Plan entire features from GitHub issues.",[46,1049,1051],{"id":1050},"the-generators-focused-role-in-the-3-core-agent-harness","The Generator's Focused Role in the 3-Core Agent Harness",[13,1053,941,1054,1056,1057],{},[636,1055,638],{},", the Generator executes solely on Planner outlines in isolated environments like Git worktrees, producing code, content, or artifacts while avoiding self-evaluation bias and memory bloat. It uses tools sparingly (search, calc) for clean, iterable outputs. Archon V3's Markdown commands make this reusable across tasks.",[17,1058,1059],{},[20,1060,108],{"href":105,"ariaDescribedBy":1061,"dataFootnoteRef":8,"id":117},[24],[13,1063,1064],{},"Core loop: Think, tool if needed, observe, repeat—but single-purpose. No chit-chat; output to files for persistence.",[13,1066,1067],{},"Contrast multi-role agents: Combining generation and eval leads to overconfidence. Generators praise mediocre work; separation forces honesty.",[13,1069,1070],{},"Examples abound. For newsletters: Planner outlines sections, Generator drafts in Markdown. For code: Implement features in fresh worktrees, committing artifacts.",[13,1072,1073,1074,1079],{},"This ties to spec-centric flows like SDD or SpecKit—Generator translates living specs without drift.",[17,1075,1076],{},[20,1077,146],{"href":143,"ariaDescribedBy":1078,"dataFootnoteRef":8,"id":155},[24]," No shared memory; pass via files.",[13,1081,1082],{},"In practice, isolation prevents hallucinations from long contexts. I've generated 50+ PRs weekly this way—far beyond chat-based tinkering.",[46,1084,1086],{"id":1085},"why-a-separate-evaluator-delivers-production-quality","Why a Separate Evaluator Delivers Production Quality",[13,1088,1089,1090,1092,1093],{},"A dedicated Evaluator in the ",[636,1091,638],{}," scores outputs adversarially against graded rubrics—like UI axes (design: 25%, originality: 25%, craft: 25%, functionality: 25%)—iterating until standards hit and upgrading GSD's pass\u002Ffail to nuanced critique assuming bugs exist. Anthropic stresses holistic excellence over TDD alone.",[17,1094,1095],{},[20,1096,26],{"href":22,"ariaDescribedBy":1097,"dataFootnoteRef":8,"id":1098},[24],"user-content-fnref-1-8",[13,1100,1101],{},"Rubrics weight criteria explicitly. UI example: \"Design: Modern, intuitive? Originality: Avoids gradients? Craft: Typography\u002Fspacing precise? Functionality: Playwright-tested?\"",[13,1103,1104],{},"Why separate? Generators optimize for completion, not quality—they hallucinate strengths. Evaluator simulates users, probing edge cases.",[13,1106,1107,1108],{},"Loop: Critique → Regenerate → Rescore until pass. Agent Blueprint pairs evaluator-optimizer for this.",[17,1109,1110],{},[20,1111,79],{"href":76,"ariaDescribedBy":1112,"dataFootnoteRef":8,"id":212},[24],[13,1114,1115],{},"Examples scale to subjective work. UI grading rejects cookie-cutter designs; content rubrics check voice alignment.",[46,1117,1119],{"id":1118},"essential-primitives-for-reliable-3-core-agent-harnesses","Essential Primitives for Reliable 3-Core Agent Harnesses",[13,1121,1122,1123,1126,1127,1132],{},"Production ",[636,1124,1125],{},"3-core agent harnesses"," require \"boring\" primitives like dual tool registries (207+ commands), tiered permissions (18 bash modules), JSON state persistence, token budgeting with projections, structured streaming, and verification. As detailed in Nate B. Jones's analysis of Claude's leaked code, these essentials harden demos into scalable systems handling crashes, controlling costs, and ensuring 24\u002F7 operation.",[17,1128,1129],{},[20,1130,34],{"href":31,"ariaDescribedBy":1131,"dataFootnoteRef":8,"id":162},[24]," Without them, you're prototyping, not producing.",[183,1134,1135,1143,1151,1154,1157,1160],{},[186,1136,1137,1138],{},"Dual tool registries (207+ commands)",[17,1139,1140],{},[20,1141,34],{"href":31,"ariaDescribedBy":1142,"dataFootnoteRef":8,"id":236},[24],[186,1144,1145,1146],{},"Tiered permissions (18 bash modules)",[17,1147,1148],{},[20,1149,34],{"href":31,"ariaDescribedBy":1150,"dataFootnoteRef":8,"id":247},[24],[186,1152,1153],{},"JSON state persistence",[186,1155,1156],{},"Token budgeting with projections",[186,1158,1159],{},"Structured streaming",[186,1161,1162],{},"Verification",[13,1164,1165,1166],{},"Nate B. Jones details 12 keys from the Claude code leak: Registries filter runtime tools dynamically; permissions gate risks (built-in > plugins > user-defined).",[17,1167,1168],{},[20,1169,34],{"href":31,"ariaDescribedBy":1170,"dataFootnoteRef":8,"id":354},[24],[13,1172,1173,1176],{},[636,1174,1175],{},"State Persistence",": Full session JSON (messages, tokens, config) for crash recovery.",[13,1178,1179],{},"Example:",[264,1181,1185],{"className":1182,"code":1183,"language":1184,"meta":8,"style":8},"language-json shiki shiki-themes github-light github-dark","{\n  \"session\": {\n    \"messages\": [...],\n    \"tokens_used\": 12500,\n    \"config\": {\"model\": \"claude-3.5-sonnet\"}\n  }\n}\n","json",[270,1186,1187,1192,1200,1214,1227,1246,1251],{"__ignoreMap":8},[273,1188,1189],{"class":275,"line":276},[273,1190,1191],{"class":283},"{\n",[273,1193,1194,1197],{"class":275,"line":287},[273,1195,1196],{"class":1024},"  \"session\"",[273,1198,1199],{"class":283},": {\n",[273,1201,1202,1205,1207,1211],{"class":275,"line":303},[273,1203,1204],{"class":1024},"    \"messages\"",[273,1206,730],{"class":283},[273,1208,1210],{"class":1209},"s7hpK","...",[273,1212,1213],{"class":283},"],\n",[273,1215,1216,1219,1221,1224],{"class":275,"line":314},[273,1217,1218],{"class":1024},"    \"tokens_used\"",[273,1220,296],{"class":283},[273,1222,1223],{"class":1024},"12500",[273,1225,1226],{"class":283},",\n",[273,1228,1229,1232,1235,1238,1240,1243],{"class":275,"line":724},[273,1230,1231],{"class":1024},"    \"config\"",[273,1233,1234],{"class":283},": {",[273,1236,1237],{"class":1024},"\"model\"",[273,1239,296],{"class":283},[273,1241,1242],{"class":299},"\"claude-3.5-sonnet\"",[273,1244,1245],{"class":283},"}\n",[273,1247,1248],{"class":275,"line":739},[273,1249,1250],{"class":283},"  }\n",[273,1252,1253],{"class":275,"line":752},[273,1254,1245],{"class":283},[13,1256,1257],{},"Save on every tool call; resume seamlessly.",[13,1259,1260,1263],{},[636,1261,1262],{},"Budgeting",": Hard limits + projections. Pre-compute: \"This task forecasts 50k input + 20k output tokens.\"",[264,1265,1269],{"className":1266,"code":1267,"language":1268,"meta":8,"style":8},"language-python shiki shiki-themes github-light github-dark","def project_tokens(prompt_len, steps=10):\n    return prompt_len * 1.2 * steps  # Conservative multiplier\nif project_tokens(len(planner_outline)) > 100000:\n    abort(\"Budget exceeded\")\n","python",[270,1270,1271,1276,1281,1286],{"__ignoreMap":8},[273,1272,1273],{"class":275,"line":276},[273,1274,1275],{},"def project_tokens(prompt_len, steps=10):\n",[273,1277,1278],{"class":275,"line":287},[273,1279,1280],{},"    return prompt_len * 1.2 * steps  # Conservative multiplier\n",[273,1282,1283],{"class":275,"line":303},[273,1284,1285],{},"if project_tokens(len(planner_outline)) > 100000:\n",[273,1287,1288],{"class":275,"line":314},[273,1289,1290],{},"    abort(\"Budget exceeded\")\n",[13,1292,1293],{},"I've seen major cost savings with budgeting alone on weekly PR batches.",[13,1295,1296,1299,1300],{},[636,1297,1298],{},"Observability",": Typed events\u002Flogs. Stream: ",[270,1301,1302],{},"{\"event\": \"tool_call\", \"tool\": \"git\", \"result\": \"committed\"}",[13,1304,1305,1308],{},[636,1306,1307],{},"Tool Registry",": Metadata-driven:",[264,1310,1312],{"className":266,"code":1311,"language":268,"meta":8,"style":8},"tools:\n  git:\n    desc: \"Git operations\"\n    perms: \"built-in\"\n    args: [\"commit\", \"push\"]\n  playwright:\n    desc: \"UI tests\"\n    perms: \"plugin\"\n",[270,1313,1314,1320,1327,1337,1347,1364,1371,1380],{"__ignoreMap":8},[273,1315,1316,1318],{"class":275,"line":276},[273,1317,280],{"class":279},[273,1319,284],{"class":283},[273,1321,1322,1325],{"class":275,"line":287},[273,1323,1324],{"class":279},"  git",[273,1326,284],{"class":283},[273,1328,1329,1332,1334],{"class":275,"line":303},[273,1330,1331],{"class":279},"    desc",[273,1333,296],{"class":283},[273,1335,1336],{"class":299},"\"Git operations\"\n",[273,1338,1339,1342,1344],{"class":275,"line":314},[273,1340,1341],{"class":279},"    perms",[273,1343,296],{"class":283},[273,1345,1346],{"class":299},"\"built-in\"\n",[273,1348,1349,1352,1354,1357,1359,1362],{"class":275,"line":724},[273,1350,1351],{"class":279},"    args",[273,1353,730],{"class":283},[273,1355,1356],{"class":299},"\"commit\"",[273,1358,801],{"class":283},[273,1360,1361],{"class":299},"\"push\"",[273,1363,736],{"class":283},[273,1365,1366,1369],{"class":275,"line":739},[273,1367,1368],{"class":279},"  playwright",[273,1370,284],{"class":283},[273,1372,1373,1375,1377],{"class":275,"line":752},[273,1374,1331],{"class":279},[273,1376,296],{"class":283},[273,1378,1379],{"class":299},"\"UI tests\"\n",[273,1381,1382,1384,1386],{"class":275,"line":763},[273,1383,1341],{"class":279},[273,1385,296],{"class":283},[273,1387,1388],{"class":299},"\"plugin\"\n",[13,1390,1391,1394,1395,1398],{},[636,1392,1393],{},"Permissions",": State objects—no world access. ",[270,1396,1397],{},"permissions: {\"bash\": [\"ls\", \"cat\"], \"deny\": [\"rm -rf\"]}",[17,1399,1400],{},[20,1401,34],{"href":31,"ariaDescribedBy":1402,"dataFootnoteRef":8,"id":1403},[24],"user-content-fnref-2-7",[13,1405,1406,1408],{},[636,1407,1162],{},": Guardrail tests post-output. Archon V3 adds pre\u002Fpost-tool hooks:",[264,1410,1412],{"className":266,"code":1411,"language":268,"meta":8,"style":8},"hooks:\n  pre-tool: \"validate_args(tool, args)\"\n  post-tool: \"verify_output_schema(result)\"\n",[270,1413,1414,1421,1431],{"__ignoreMap":8},[273,1415,1416,1419],{"class":275,"line":276},[273,1417,1418],{"class":279},"hooks",[273,1420,284],{"class":283},[273,1422,1423,1426,1428],{"class":275,"line":287},[273,1424,1425],{"class":279},"  pre-tool",[273,1427,296],{"class":283},[273,1429,1430],{"class":299},"\"validate_args(tool, args)\"\n",[273,1432,1433,1436,1438],{"class":275,"line":303},[273,1434,1435],{"class":279},"  post-tool",[273,1437,296],{"class":283},[273,1439,1440],{"class":299},"\"verify_output_schema(result)\"\n",[13,1442,1443,1446,1447],{},[636,1444,1445],{},"Agent Typing",": Modes like explore\u002Fplan\u002Fverify; permissions evolve with state.",[17,1448,1449],{},[20,1450,34],{"href":31,"ariaDescribedBy":1451,"dataFootnoteRef":8,"id":1452},[24],"user-content-fnref-2-8",[13,1454,1455],{},"Dynamic pools swap tools mid-run. Skip these, long tasks fail frequently—primitives ensure 24\u002F7 operation.",[46,1457,1459],{"id":1458},"building-and-scaling-your-3-core-agent-harness","Building and Scaling Your 3-Core Agent Harness",[13,1461,1462,1463,1469,1475],{},"Assemble a 3-core agent harness via YAML\u002FDAGs (Archon V3), Claude teams, or SpecKit staging: Define each core with Role+Goal+Tools+Rules+Output, layer primitives, master single-agent first, then scale multi for skills.",[17,1464,1465],{},[20,1466,79],{"href":76,"ariaDescribedBy":1467,"dataFootnoteRef":8,"id":1468},[24],"user-content-fnref-3-5",[17,1470,1471],{},[20,1472,108],{"href":105,"ariaDescribedBy":1473,"dataFootnoteRef":8,"id":1474},[24],"user-content-fnref-4-3"," GitHub Actions trigger workflows; minimal tools\u002Fmemory.",[13,1477,1478],{},"Steps:",[401,1480,1481,1491,1497],{},[186,1482,1483,1486,1487,1490],{},[636,1484,1485],{},"Planner Prompt",": \"Role: Product lead. Goal: ",[273,1488,1489],{},"Task",". Tools: Search. Rules: High-level only. Output: Markdown outline.\"",[186,1492,1493,1496],{},[636,1494,1495],{},"Generator",": \"Role: Engineer. Goal: Implement outline. Tools: Git, calc. Rules: Isolated worktree. Output: Committed artifacts.\"",[186,1498,1499,1502],{},[636,1500,1501],{},"Evaluator",": \"Role: QA adversary. Goal: Score vs rubric. Tools: Playwright. Rules: Assume flaws. Output: Scores + fixes.\"",[13,1504,1505],{},"Add primitives: Token budgeter, JSON state.",[13,1507,1508,1509],{},"Test on messy inputs. Ready starters: GSD + rubrics; Archon builtins (fix-issue, idea-to-PR).",[17,1510,1511],{},[20,1512,108],{"href":105,"ariaDescribedBy":1513,"dataFootnoteRef":8,"id":1514},[24],"user-content-fnref-4-4",[13,1516,1517],{},"Scaling: Orchestrators route; parallelism via DAGs.",[13,1519,1520],{},"Harness over frameworks—I've shipped 10x faster.",[46,1522,1524],{"id":1523},"trade-offs-misconceptions-and-when-to-skip-agents","Trade-offs, Misconceptions, and When to Skip Agents",[13,1526,1527,1528],{},"The 3-core agent harness swaps framework \"safety\" for speed but requires primitive investment; it costs more upfront yet runs leaner long-term, with budgeting taming expenses. Misconceptions: Micro-tasks aren't needed (LLMs obsolete them); always multi-agent (master single first). Skip for deterministic chaining\u002Frouting—use Anthropic's 5 workflows.",[17,1529,1530],{},[20,1531,79],{"href":76,"ariaDescribedBy":1532,"dataFootnoteRef":8,"id":1533},[24],"user-content-fnref-3-6",[13,1535,1536],{},"Trade-offs:",[183,1538,1539,1545,1551],{},[186,1540,1541,1544],{},[636,1542,1543],{},"Build Time",": Primitives take days vs. framework npm install.",[186,1546,1547,1550],{},[636,1548,1549],{},"Eval Subjectivity",": Rubrics need tuning for creative work.",[186,1552,1553,1556,1557],{},[636,1554,1555],{},"Costs",": Long tasks hit tokens—mitigate with projections.",[17,1558,1559],{},[20,1560,34],{"href":31,"ariaDescribedBy":1561,"dataFootnoteRef":8,"id":1562},[24],"user-content-fnref-2-9",[13,1564,1565,1566,1572],{},"Risks: No permissions = demo, not product.",[17,1567,1568],{},[20,1569,34],{"href":31,"ariaDescribedBy":1570,"dataFootnoteRef":8,"id":1571},[24],"user-content-fnref-2-10"," Frameworks aren't future-proof; LLMs evolve.",[13,1574,1575],{},"When to skip: Simple transforms? Chain prompts. Agents shine on open-ended horizons.",[13,1577,1578],{},"Open: Rubric standards? This works today—iterate your own.",[13,1580,1581],{},"Pick a stalled repo project. Strip any agent setup to these three roles using Role+Goal+Tools+Rules+Output. Add token budgeting as your first primitive. Run on one feature, measure output quality and costs before scaling—you'll ship faster than with frameworks.",[391,1583,1585,1588],{"className":1584,"dataFootnotes":8},[394],[46,1586,399],{"className":1587,"id":24},[398],[401,1589,1590,1634,1695,1730,1755],{},[186,1591,1592,1593,414,1596,414,1601,414,1606,414,1611,414,1616,414,1621,414,1626],{"id":405},"AI LABS, \"Anthropic: Agent Harnesses Need Only 3 Core Agents\" (YouTube summary). ",[20,1594,413],{"href":409,"ariaLabel":410,"className":1595,"dataFootnoteBackref":8},[412],[20,1597,413,1599],{"href":417,"ariaLabel":418,"className":1598,"dataFootnoteBackref":8},[412],[17,1600,34],{},[20,1602,413,1604],{"href":424,"ariaLabel":425,"className":1603,"dataFootnoteBackref":8},[412],[17,1605,79],{},[20,1607,413,1609],{"href":431,"ariaLabel":432,"className":1608,"dataFootnoteBackref":8},[412],[17,1610,108],{},[20,1612,413,1614],{"href":438,"ariaLabel":439,"className":1613,"dataFootnoteBackref":8},[412],[17,1615,146],{},[20,1617,413,1619],{"href":445,"ariaLabel":446,"className":1618,"dataFootnoteBackref":8},[412],[17,1620,343],{},[20,1622,413,1624],{"href":452,"ariaLabel":453,"className":1623,"dataFootnoteBackref":8},[412],[17,1625,457],{},[20,1627,413,1631],{"href":1628,"ariaLabel":1629,"className":1630,"dataFootnoteBackref":8},"#user-content-fnref-1-8","Back to reference 1-8",[412],[17,1632,1633],{},"8",[186,1635,1636,1637,414,1640,414,1645,414,1650,414,1655,414,1660,414,1665,414,1672,414,1679,414,1687],{"id":460},"Nate B. Jones, \"Claude Code Leak: 12 Primitives for Production Agents\" (YouTube). ",[20,1638,413],{"href":464,"ariaLabel":465,"className":1639,"dataFootnoteBackref":8},[412],[20,1641,413,1643],{"href":469,"ariaLabel":470,"className":1642,"dataFootnoteBackref":8},[412],[17,1644,34],{},[20,1646,413,1648],{"href":476,"ariaLabel":477,"className":1647,"dataFootnoteBackref":8},[412],[17,1649,79],{},[20,1651,413,1653],{"href":483,"ariaLabel":484,"className":1652,"dataFootnoteBackref":8},[412],[17,1654,108],{},[20,1656,413,1658],{"href":490,"ariaLabel":491,"className":1657,"dataFootnoteBackref":8},[412],[17,1659,146],{},[20,1661,413,1663],{"href":497,"ariaLabel":498,"className":1662,"dataFootnoteBackref":8},[412],[17,1664,343],{},[20,1666,413,1670],{"href":1667,"ariaLabel":1668,"className":1669,"dataFootnoteBackref":8},"#user-content-fnref-2-7","Back to reference 2-7",[412],[17,1671,457],{},[20,1673,413,1677],{"href":1674,"ariaLabel":1675,"className":1676,"dataFootnoteBackref":8},"#user-content-fnref-2-8","Back to reference 2-8",[412],[17,1678,1633],{},[20,1680,413,1684],{"href":1681,"ariaLabel":1682,"className":1683,"dataFootnoteBackref":8},"#user-content-fnref-2-9","Back to reference 2-9",[412],[17,1685,1686],{},"9",[20,1688,413,1692],{"href":1689,"ariaLabel":1690,"className":1691,"dataFootnoteBackref":8},"#user-content-fnref-2-10","Back to reference 2-10",[412],[17,1693,1694],{},"10",[186,1696,1697,1698,414,1701,414,1706,414,1711,414,1716,414,1723],{"id":504},"Lukas Margerie, \"Agent Blueprint: Role + Goal + Tools + Rules + Output\" (YouTube). ",[20,1699,413],{"href":508,"ariaLabel":509,"className":1700,"dataFootnoteBackref":8},[412],[20,1702,413,1704],{"href":513,"ariaLabel":514,"className":1703,"dataFootnoteBackref":8},[412],[17,1705,34],{},[20,1707,413,1709],{"href":520,"ariaLabel":521,"className":1708,"dataFootnoteBackref":8},[412],[17,1710,79],{},[20,1712,413,1714],{"href":527,"ariaLabel":528,"className":1713,"dataFootnoteBackref":8},[412],[17,1715,108],{},[20,1717,413,1721],{"href":1718,"ariaLabel":1719,"className":1720,"dataFootnoteBackref":8},"#user-content-fnref-3-5","Back to reference 3-5",[412],[17,1722,146],{},[20,1724,413,1728],{"href":1725,"ariaLabel":1726,"className":1727,"dataFootnoteBackref":8},"#user-content-fnref-3-6","Back to reference 3-6",[412],[17,1729,343],{},[186,1731,1732,1733,414,1736,414,1741,414,1748],{"id":534},"DIY Smart Code, \"Archon V3: YAML Harnesses for AI Coding Agents\" (YouTube). ",[20,1734,413],{"href":538,"ariaLabel":539,"className":1735,"dataFootnoteBackref":8},[412],[20,1737,413,1739],{"href":543,"ariaLabel":544,"className":1738,"dataFootnoteBackref":8},[412],[17,1740,34],{},[20,1742,413,1746],{"href":1743,"ariaLabel":1744,"className":1745,"dataFootnoteBackref":8},"#user-content-fnref-4-3","Back to reference 4-3",[412],[17,1747,79],{},[20,1749,413,1753],{"href":1750,"ariaLabel":1751,"className":1752,"dataFootnoteBackref":8},"#user-content-fnref-4-4","Back to reference 4-4",[412],[17,1754,108],{},[186,1756,1757,1758,414,1761],{"id":550},"Level Up Coding, \"SDD Makes Specs the Single Source of Truth via AI Agents.\" ",[20,1759,413],{"href":554,"ariaLabel":555,"className":1760,"dataFootnoteBackref":8},[412],[20,1762,413,1764],{"href":559,"ariaLabel":560,"className":1763,"dataFootnoteBackref":8},[412],[17,1765,34],{},[588,1767,1768],{},"html pre.shiki code .s9eBZ, html code.shiki .s9eBZ{--shiki-default:#22863A;--shiki-dark:#85E89D}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .s7hpK, html code.shiki .s7hpK{--shiki-default:#B31D28;--shiki-default-font-style:italic;--shiki-dark:#FDAEB7;--shiki-dark-font-style:italic}",{"title":8,"searchDepth":287,"depth":287,"links":1770},[1771,1772,1773,1774,1775,1776,1777,1778],{"id":896,"depth":287,"text":897},{"id":937,"depth":287,"text":938},{"id":1050,"depth":287,"text":1051},{"id":1085,"depth":287,"text":1086},{"id":1118,"depth":287,"text":1119},{"id":1458,"depth":287,"text":1459},{"id":1523,"depth":287,"text":1524},{"id":24,"depth":287,"text":399},"ai-llms","Production AI agent systems thrive with a 3-core agent harness—Planner for high-level outlines, Generator for implementation, Evaluator for adversarial critique using graded rubrics—stripping 90% bloated framework overhead per Anthropic tests.1 This leverages modern LLMs' 1M+ token windows and coherence for reliable, scalable outputs on long-horizon tasks.",[638,612,1782,1783,1784,1785,607,1786],"planner generator evaluator","agent frameworks","Claude Opus","AI agent primitives","minimalist agents",{},"Learn why production AI agents need just Planner, Generator, and Evaluator—not frameworks. Anthropic tests show 90% framework bloat is obsolete. Build scalable harnesses with primitives for reliable outputs in code, UI, and content.","3-Core Agent Harness for Production AI","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-article",[],{"intro":1793,"why-production-agent-systems-need-a-3-core-agent-harness":897,"the-planners-role-in-a-3-core-agent-harness":938,"the-generators-focused-role-in-the-3-core-agent-harness":1051,"why-a-separate-evaluator-delivers-production-quality":1086,"essential-primitives-for-reliable-3-core-agent-harnesses":1119,"building-and-scaling-your-3-core-agent-harness":1459,"trade-offs-misconceptions-and-when-to-skip-agents":1524},"Introduction",{"title":628,"description":1780},{"loc":1790},"3-core-agent-harness-planner-generator-evaluator-production-ai-agents","articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-article","pEqxcDCwoMVirkXg2PYcYkpR4uAFqYg3dO1nMbJGObo",{"id":1800,"title":1801,"ai_summary":1802,"author":8,"body":1803,"category":1779,"date_modified":601,"description":2395,"excerpt":8,"extension":603,"faq":601,"keywords":2396,"meta":2401,"meta_description":2402,"meta_title":2403,"navigation":616,"path":2404,"primary_keyword":1867,"published_at":601,"related_posts":2405,"sections":2406,"seo":2408,"sitemap":2409,"slug":1796,"status":623,"stem":2410,"__hash__":2411},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-baseline-post-merge-article.md","The 3-Core Agent Harness: Planner, Generator, Evaluator for Reliable Production AI Agents","Production AI agent systems require a 3-core agent harness—Planner for task decomposition, Generator for execution, Evaluator for unbiased verification—to fix single-agent issues like underspecification, self-evaluation bias, and context limits, ensuring reliable outputs on complex, long-running tasks.",{"type":10,"value":1804,"toc":2385},[1805,1811,1814,1822,1826,1829,1836,1848,1855,1858,1862,1869,1878,1881,1899,1906,1909,1913,1916,1919,1922,1942,1949,1952,2042,2050,2054,2057,2064,2067,2087,2095,2103,2106,2110,2116,2121,2124,2129,2132,2143,2148,2155,2223,2226,2230,2236,2248,2251,2276,2284,2287,2291,2296,2306,2312,2323,2326,2346,2349,2353,2359,2362,2379,2382],[13,1806,1807,1808,1810],{},"Production AI agent systems demand a ",[636,1809,638],{},"—Planner for task decomposition, Generator for execution, Evaluator for unbiased verification—because single agents suffer from underspecification, self-bias, and context limits, delivering unreliable outputs on complex tasks. (38 words)",[13,1812,1813],{},"Picture this: You're a developer tasked with building a full-stack feature using a single-LLM agent. It spits out a toy dashboard—missing auth, no error handling, half-baked UI. Hours vanish fixing hallucinations and gaps. Human teams avoid this by separating planning, coding, and testing. Agents need the same discipline.",[13,1815,1816,1817,1821],{},"Anthropic's research nails why single agents crumble in production: they underscope tasks, rubber-stamp their own bugs, and panic at context limits.",[17,1818,1819],{},[273,1820,26],{}," This harness forms a feedback loop, like a GAN for code generation, where agents critique each other. Builders shipping AI products face demo-to-production chasms today—skyrocketing debug time, token costs eating margins. We'll cover why singles fail, the harness blueprint, core roles, scaling patterns, framework limits, prod fixes, and your first build.",[46,1823,1825],{"id":1824},"why-single-agent-systems-fail-production-tasks","Why Single-Agent Systems Fail Production Tasks",[13,1827,1828],{},"Single-agent LLM systems fail production tasks due to underspecification—where vague prompts lead to simplified outputs—their inability to self-critique without bias, context window anxiety that rushes incomplete work, and lack of structured planning for multi-step problems, as shown in Anthropic's analysis of real-world agent breakdowns. (62 words)",[13,1830,1831,1832],{},"High-level prompts sound clear to humans but trip up LLMs. Take \"build a user dashboard\": a single agent might deliver a static HTML mockup, skipping backend integration or scalability. Engr. Mejba Ahmed documented this in his Anthropic harness experiments: agents consistently produced \"toy versions,\" ignoring full specs like persistence or security.",[17,1833,1834],{},[273,1835,34],{},[13,1837,1838,1839,1843,1844],{},"Self-evaluation bias compounds the mess. LLMs generating code often approve it uncritically, spotting syntax nits but missing logic flaws. Nurunnubi Talukder puts it bluntly: \"Having the same agent generate and then judge its own output just doesn’t really work. Decoupling those roles seems pretty key.\"",[17,1840,1841],{},[273,1842,79],{}," Studies confirm LLMs inflate their scores by 20-30% on flawed work.",[17,1845,1846],{},[273,1847,108],{},[13,1849,1850,1851],{},"Context anxiety hits long tasks hardest. As tokens pile up, agents truncate reasoning or output prematurely. Atal Upadhyay notes workflow failures here: no native decomposition means no handling for sprints or state.",[17,1852,1853],{},[273,1854,146],{},[13,1856,1857],{},"Humans sidestep this with role separation—PM plans, dev builds, QA tests. Single agents mash them, breeding errors. Production demands the split.",[46,1859,1861],{"id":1860},"what-is-the-3-core-agent-harness","What Is the 3-Core Agent Harness?",[13,1863,1864,1865,1868],{},"The 3-core ",[636,1866,1867],{},"agent harness"," is an architecture with specialized Planner, Generator, and Evaluator agents that decomposes high-level goals into specs, executes them iteratively, and verifies outputs objectively, overcoming single-agent limits for reliable production AI applications like full-stack development or long-running automations. This setup forms a tight feedback loop: Planner sets contracts, Generator builds sprints, Evaluator scores and iterates—mimicking GANs but for tasks. (78 words)",[13,1870,1871,1872,1877],{},"It's not just prompts; it's infrastructure. Anthropic's 2026 paper on long-running development formalized it, with Zylos Research expanding on patterns.",[17,1873,1874,1875],{},"[6]",[273,1876,457],{}," Separation kills bias: Generator can't fudge evals.",[13,1879,1880],{},"The loop runs like this:",[183,1882,1883,1889,1894],{},[186,1884,1885,1888],{},[636,1886,1887],{},"Planner"," outputs sprint specs with \"done\" criteria.",[186,1890,1891,1893],{},[636,1892,1495],{}," executes one sprint at a time.",[186,1895,1896,1898],{},[636,1897,1501],{}," tests via tools (e.g., Playwright for UI), scores, loops back if needed.",[13,1900,1901,1902],{},"This enables tool use—evaluators run browsers, not just read code. Upadhyay calls it shifting from \"reading about the work\" to \"experiencing it.\"",[17,1903,1904],{},[273,1905,146],{},[13,1907,1908],{},"Why now? Demos hide flaws; production exposes them under load. Harnesses bridge that, cutting debug time I've seen drop 40% in my builds.",[46,1910,1912],{"id":1911},"the-planner-decomposing-tasks-into-actionable-specs","The Planner: Decomposing Tasks into Actionable Specs",[13,1914,1915],{},"The Planner agent transforms vague user prompts into detailed product specs by breaking tasks into sequenced sub-tasks or sprints, defining \"done\" criteria, and tracking shared state, preventing underscoping and enabling ambitious, complete outputs without over-specifying implementation details. (50 words)",[13,1917,1918],{},"Input a goal like \"build a task manager app.\" Output: Sprint 1 (auth + DB schema, criteria: JWT login passes, data persists); Sprint 2 (UI CRUD, criteria: E2E tests green). No code dictums—just contracts.",[13,1920,1921],{},"Principles keep it sharp:",[183,1923,1924,1930,1936],{},[186,1925,1926,1929],{},[636,1927,1928],{},"Ambitious scope",": Aim high; partial wins beat toys.",[186,1931,1932,1935],{},[636,1933,1934],{},"Implementation agnostic",": Specs focus outcomes, not stacks.",[186,1937,1938,1941],{},[636,1939,1940],{},"State awareness",": Track files, progress.",[13,1943,1944,1945],{},"Atal Upadhyay's example: Decomposing a web scraper into auth, crawl, parse sprints yielded 2x completeness vs. singles.",[17,1946,1947],{},[273,1948,146],{},[13,1950,1951],{},"Pseudocode for a Planner prompt:",[264,1953,1955],{"className":1266,"code":1954,"language":1268,"meta":8,"style":8},"prompt = \"\"\"\nYou are the Planner. Given a high-level goal: {goal}\n\nOutput JSON:\n{{\n  \"sprints\": [\n    {{\n      \"id\": 1,\n      \"description\": \"Detailed sub-task\",\n      \"done_criteria\": [\"Testable outcomes\"],\n      \"state_files\": [\"shared.json\"]\n    }}\n  ],\n  \"total_sprints\": N\n}}\nAmbitious but realistic. No implementation details.\n\"\"\"\n",[270,1956,1957,1962,1967,1972,1977,1982,1987,1992,1997,2002,2007,2012,2017,2022,2027,2032,2037],{"__ignoreMap":8},[273,1958,1959],{"class":275,"line":276},[273,1960,1961],{},"prompt = \"\"\"\n",[273,1963,1964],{"class":275,"line":287},[273,1965,1966],{},"You are the Planner. Given a high-level goal: {goal}\n",[273,1968,1969],{"class":275,"line":303},[273,1970,1971],{"emptyLinePlaceholder":616},"\n",[273,1973,1974],{"class":275,"line":314},[273,1975,1976],{},"Output JSON:\n",[273,1978,1979],{"class":275,"line":724},[273,1980,1981],{},"{{\n",[273,1983,1984],{"class":275,"line":739},[273,1985,1986],{},"  \"sprints\": [\n",[273,1988,1989],{"class":275,"line":752},[273,1990,1991],{},"    {{\n",[273,1993,1994],{"class":275,"line":763},[273,1995,1996],{},"      \"id\": 1,\n",[273,1998,1999],{"class":275,"line":771},[273,2000,2001],{},"      \"description\": \"Detailed sub-task\",\n",[273,2003,2004],{"class":275,"line":781},[273,2005,2006],{},"      \"done_criteria\": [\"Testable outcomes\"],\n",[273,2008,2009],{"class":275,"line":791},[273,2010,2011],{},"      \"state_files\": [\"shared.json\"]\n",[273,2013,2014],{"class":275,"line":809},[273,2015,2016],{},"    }}\n",[273,2018,2019],{"class":275,"line":821},[273,2020,2021],{},"  ],\n",[273,2023,2024],{"class":275,"line":831},[273,2025,2026],{},"  \"total_sprints\": N\n",[273,2028,2029],{"class":275,"line":839},[273,2030,2031],{},"}}\n",[273,2033,2034],{"class":275,"line":849},[273,2035,2036],{},"Ambitious but realistic. No implementation details.\n",[273,2038,2039],{"class":275,"line":859},[273,2040,2041],{},"\"\"\"\n",[13,2043,2044,2045,2049],{},"Avijit M nails the shift: \"We’ve been building AI apps the wrong way. Real-world problems don’t work like one prompt → one output.\"",[17,2046,2047],{},[273,2048,1633],{}," Planners force clarity.",[46,2051,2053],{"id":2052},"generator-and-evaluator-execution-meets-verification","Generator and Evaluator: Execution Meets Verification",[13,2055,2056],{},"The Generator executes Planner specs within sprints, focusing purely on building code or content, while the separate Evaluator tests outputs using tools like browser automation or test suites for objective scoring, eliminating self-bias and ensuring functional results through feedback loops. (52 words)",[13,2058,2059,2060],{},"Generator takes a sprint contract: \"Build auth endpoint. Criteria: POST \u002Flogin returns JWT, stores user.\" It outputs code, negotiating if specs shift. Mejba Ahmed's builds showed generators hit 85% sprint success when focused.",[17,2061,2062],{},[273,2063,34],{},[13,2065,2066],{},"Evaluator is the star. Armed with tools:",[183,2068,2069,2075,2081],{},[186,2070,2071,2074],{},[636,2072,2073],{},"Playwright"," for UI: Launches app, clicks flows.",[186,2076,2077,2080],{},[636,2078,2079],{},"Linters\u002Ftests"," for code: Runs suite, flags fails.",[186,2082,2083,2086],{},[636,2084,2085],{},"Rubrics"," for subjective: Scores 1-10 on criteria.",[13,2088,2089,2090,2094],{},"No more \"nod and pretend.\" Upadhyay: \"Text-only eval is insufficient... Use tools to experience the work.\"",[17,2091,2092],{},[273,2093,146],{}," Scores feed back: \u003C7? Regenerate.",[13,2096,2097,2098,2102],{},"Contrast self-eval: Ahmed found separate critics caught 3x more bugs.",[17,2099,2100],{},[273,2101,34],{}," In harnesses, Generator builds blindly; Evaluator judges coldly. Loops cap at budgets—e.g., 3 tries per sprint.",[13,2104,2105],{},"This duo powers the harness core. I've refactored agent fails this way; quality jumps.",[46,2107,2109],{"id":2108},"agent-harness-patterns-for-scaling-complexity","Agent Harness Patterns for Scaling Complexity",[13,2111,2112,2115],{},[636,2113,2114],{},"Agent harness"," patterns scale from Simple Loop (single agent for quick validations) to Generator-Evaluator pairs for subjective quality, up to full 3-core harness for complex tasks, matching architecture to task needs for optimal cost and reliability. Pick by complexity: low for math checks, full for apps. (56 words)",[13,2117,2118],{},[636,2119,2120],{},"Pattern A: Simple Loop",[13,2122,2123],{},"One agent loops with hard checks (linters). Low tokens, fast for scripts. Cost: ~10% of full.",[13,2125,2126],{},[636,2127,2128],{},"Pattern B: Gen-Eval Pair",[13,2130,2131],{},"Two agents: Gen builds, Eval rubrics\u002Fscores. Mid-tasks like content. Rubric example:",[183,2133,2134,2137,2140],{},[186,2135,2136],{},"Functionality: 40%",[186,2138,2139],{},"Best practices: 30%",[186,2141,2142],{},"Edge cases: 30%",[13,2144,2145],{},[636,2146,2147],{},"Pattern C: Full PGE (Planner + Gen + Eval)",[13,2149,2150,2151],{},"Sprints for stacks. Upadhyay's guide: Use for >5 steps.",[17,2152,2153],{},[273,2154,146],{},[2156,2157,2158,2177],"table",{},[2159,2160,2161],"thead",{},[2162,2163,2164,2168,2171,2174],"tr",{},[2165,2166,2167],"th",{},"Pattern",[2165,2169,2170],{},"Task Fit",[2165,2172,2173],{},"Cost Multiplier",[2165,2175,2176],{},"Reliability",[2178,2179,2180,2195,2209],"tbody",{},[2162,2181,2182,2186,2189,2192],{},[2183,2184,2185],"td",{},"Simple Loop",[2183,2187,2188],{},"Short, objective",[2183,2190,2191],{},"1x",[2183,2193,2194],{},"High for basics",[2162,2196,2197,2200,2203,2206],{},[2183,2198,2199],{},"Gen-Eval",[2183,2201,2202],{},"Mid, subjective",[2183,2204,2205],{},"2-3x",[2183,2207,2208],{},"Good",[2162,2210,2211,2214,2217,2220],{},[2183,2212,2213],{},"Full Harness",[2183,2215,2216],{},"Complex, long",[2183,2218,2219],{},"5x+",[2183,2221,2222],{},"Production-grade",[13,2224,2225],{},"Match to needs—don't overengineer.",[46,2227,2229],{"id":2228},"why-agent-frameworks-arent-enough","Why Agent Frameworks Aren't Enough",[13,2231,2232,2233,2235],{},"Agent frameworks like LangChain, CrewAI, and AutoGen provide orchestration tools but fall short without custom ",[636,2234,1867],{}," design, ignoring data governance gaps (80% of impl time on data eng) and production issues like hallucinations under load. They're Lego bricks; harness is the blueprint. (54 words)",[13,2237,2238,2239,2243,2244],{},"Rasa's 2026 review scores them on readiness: LangGraph strong on graphs, CrewAI on crews, but all demo-shine, prod-crack.",[17,2240,2241],{},[273,2242,1686],{}," Airbyte: 80% time on data pipelines, not config.",[17,2245,2246],{},[273,2247,1694],{},[13,2249,2250],{},"Gaps:",[183,2252,2253,2264,2270],{},[186,2254,2255,2258,2259],{},[636,2256,2257],{},"Data trust",": Emily Winks (Atlan): \"Agents assume trustworthy data... It's a governance problem.\"",[17,2260,2261],{},[273,2262,2263],{},"11",[186,2265,2266,2269],{},[636,2267,2268],{},"Explainability",": Black-box flows fail audits.",[186,2271,2272,2275],{},[636,2273,2274],{},"Load hallucinations",": Demos low-traffic; prod spikes break.",[13,2277,2278,2279],{},"Frameworks orchestrate; harnesses govern. Build custom atop them—LangGraph for PGE loops shines. Future: MCP standards for interoperability.",[17,2280,2281],{},[273,2282,2283],{},"12",[13,2285,2286],{},"Harness engineering trumps model power now.",[46,2288,2290],{"id":2289},"production-challenges-and-fixes-for-agent-harnesses","Production Challenges and Fixes for Agent Harnesses",[13,2292,1122,2293,2295],{},[636,2294,1867],{},"es face high costs from token usage, debugging multi-agent interactions, and data governance issues, fixed by budget enforcement, modular monitoring, tool integration, and iterative \"build for deletion\" design anticipating model improvements. Costs can 5x vs. singles, but quality pays. (48 words)",[13,2297,2298,2301,2302],{},[636,2299,2300],{},"Cost control",": Track per-agent tokens; cap loops (e.g., 3 evals max). Ahmed's runs: Harness saved 20% net via fewer redos.",[17,2303,2304],{},[273,2305,34],{},[13,2307,2308,2311],{},[636,2309,2310],{},"Debugging",": Log state, comms. Modular: Swap evals without rebuild.",[13,2313,2314,2317,2318],{},[636,2315,2316],{},"Data gov",": Validate inputs—Atlan tools pre-check staleness.",[17,2319,2320],{},[273,2321,2322],{},"13",[13,2324,2325],{},"Fixes:",[183,2327,2328,2334,2340],{},[186,2329,2330,2333],{},[636,2331,2332],{},"Human-in-loop",": Approve specs, spot drifts.",[186,2335,2336,2339],{},[636,2337,2338],{},"Scalability",": Async sprints, queueing.",[186,2341,2342,2345],{},[636,2343,2344],{},"Build for deletion",": Modular for model leaps.",[13,2347,2348],{},"Misconception: Frameworks = done. No—harness first.",[46,2350,2352],{"id":2351},"building-reliable-agent-harnesses-today","Building Reliable Agent Harnesses Today",[13,2354,2355,2356,2358],{},"Specialized roles beat generalists for complex work. The ",[636,2357,1867],{},"—Planner, Generator, Evaluator—delivers production robustness as frameworks mature but orchestration lags.",[13,2360,2361],{},"Trends: Tool evals, modularity. Checklist for your first:",[401,2363,2364,2367,2370,2373,2376],{},[186,2365,2366],{},"Audit a failed single-agent task.",[186,2368,2369],{},"Sketch PGE: Sub-tasks, criteria.",[186,2371,2372],{},"Prototype in LangGraph\u002FCrewAI.",[186,2374,2375],{},"Run one sprint; log tokens.",[186,2377,2378],{},"Iterate to 80%+ eval score.",[13,2380,2381],{},"Pick a backlog flop, paper-sketch the harness (Planner subs, Eval tools), prototype in LangGraph—track costs, one sprint. Measure the lift yourself.",[588,2383,2384],{},"html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":8,"searchDepth":287,"depth":287,"links":2386},[2387,2388,2389,2390,2391,2392,2393,2394],{"id":1824,"depth":287,"text":1825},{"id":1860,"depth":287,"text":1861},{"id":1911,"depth":287,"text":1912},{"id":2052,"depth":287,"text":2053},{"id":2108,"depth":287,"text":2109},{"id":2228,"depth":287,"text":2229},{"id":2289,"depth":287,"text":2290},{"id":2351,"depth":287,"text":2352},"Production AI agent systems demand a 3-core agent harness—Planner for task decomposition, Generator for execution, Evaluator for unbiased verification—because single agents suffer from underspecification, self-bias, and context limits, delivering unreliable outputs on complex tasks. (38 words)",[1867,609,1782,2397,2398,1783,2399,2400],"AI agents production","LLM agent architecture","harness design","AI agent evaluation",{},"Build reliable production AI agents with the 3-core harness: Planner decomposes tasks, Generator executes, Evaluator verifies objectively. Overcome single-agent failures like underscoping and bias for full-stack builds and automations—patterns, trade-offs, and implementation guide.","3-Core Agent Harness for Production AI Agents","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-baseline-post-merge-article",[],{"intro":2407,"why-single-agent-systems-fail-production-tasks":1825,"what-is-the-3-core-agent-harness":1861,"the-planner-decomposing-tasks-into-actionable-specs":1912,"generator-and-evaluator-execution-meets-verification":2053,"agent-harness-patterns-for-scaling-complexity":2109,"why-agent-frameworks-arent-enough":2229,"production-challenges-and-fixes-for-agent-harnesses":2290,"building-reliable-agent-harnesses-today":2352},"(intro)",{"title":1801,"description":2395},{"loc":2404},"articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-baseline-post-merge-article","UhWs8sgZbsGRyL5uVfvs_QzI9CHX8VgqL3qnduhOEpk",{"id":2413,"title":2414,"ai_summary":2415,"author":8,"body":2416,"category":1779,"date_modified":601,"description":3205,"excerpt":8,"extension":603,"faq":601,"keywords":3206,"meta":3213,"meta_description":3214,"meta_title":3215,"navigation":616,"path":3216,"primary_keyword":607,"published_at":601,"related_posts":3217,"sections":3218,"seo":3219,"sitemap":3220,"slug":3221,"status":623,"stem":3222,"__hash__":3223},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-baseline-round2-article.md","Harness Engineering: The 3-Core System for Reliable Production AI Agents","Production AI agents succeed through harness engineering—a structured system around the model featuring a Planner for specs, Generator for implementation, and independent Evaluator for critique—outperforming solo agents by addressing generosity bias and self-evaluation failures. This 3-core harness delivers reliable outputs for coding and long tasks. (50 words)",{"type":10,"value":2417,"toc":3193},[2418,2427,2436,2439,2443,2455,2458,2461,2481,2496,2499,2503,2512,2519,2526,2529,2538,2541,2544,2548,2557,2560,2580,2589,2592,2601,2604,2608,2617,2620,2628,2631,2634,2679,2682,2685,2689,2698,2703,2709,2712,2717,2728,2731,2734,2742,2751,2754,2773,2777,2786,2792,2798,2809,2823,2826,2875,2878,2881,2884,2888,2897,2900,2908,2924,2932,2935,2938,2942,2951,2953,2964,2967,2970,3010,3013,3016,3020,3023,3026,3029,3046,3049,3190],[13,2419,2420,2421,2426],{},"Production AI agents fail at an 88% rate despite LLM advances because they rely on solo models without proper scaffolding.",[17,2422,2423],{},[20,2424,26],{"href":22,"ariaDescribedBy":2425,"dataFootnoteRef":8,"id":25},[24]," Harness engineering fixes this: build a 3-core system with a Planner for specs, Generator for code, and independent Evaluator for critique. This outperforms single agents by countering generosity bias and self-evaluation flaws (42 words).",[13,2428,2429,2430,2435],{},"You've seen the hype—agents that code apps, manage workflows, ship features. But in production, most crumble. A solo agent might spit out a game engine in 20 minutes for $9, only for core mechanics to break on first play.",[17,2431,2432],{},[20,2433,34],{"href":31,"ariaDescribedBy":2434,"dataFootnoteRef":8,"id":33},[24]," The 3-core harness, drawn from Anthropic's research, takes six hours and $200 but delivers a fully playable title with advanced features like multiplayer and AI opponents.",[13,2437,2438],{},"Why does this matter now? As an indie builder or AI dev, you're shipping products where unreliable agents mean broken features, wasted cycles, and lost trust. Harness engineering isn't theory—it's the system that turns demos into deployables. We'll break down the why, the architecture, components, results, and trade-offs, with prompts and patterns you can use today.",[46,2440,2442],{"id":2441},"what-is-harness-engineering","What Is Harness Engineering?",[13,2444,2445,2446,987,2449,2454],{},"Harness engineering is the discipline of building production-grade scaffolding around LLMs—coordination layers, specialized agents, validation loops, and session orchestration—that ensures reliable AI agent performance, far more than model choice alone. It evolved from prompt engineering to handle complex, multi-turn tasks like app development, with components like guides (feedforward) and sensors (feedback) encoding fixes for model weaknesses. Tian Pan boils it down: ",[636,2447,2448],{},"Agent = Model + Harness",[17,2450,2451],{},[20,2452,79],{"href":76,"ariaDescribedBy":2453,"dataFootnoteRef":8,"id":78},[24]," (72 words)",[13,2456,2457],{},"Prompt engineering works for single turns: stuff context, get output. But production agents run sessions—hours of back-and-forth, context resets, accumulating errors. Harness engineering orchestrates this: it sequences agent calls, injects state, validates outputs, and loops on failures.",[13,2459,2460],{},"Core principles set it apart:",[183,2462,2463,2469,2475],{},[186,2464,2465,2468],{},[636,2466,2467],{},"Specialization",": No jack-of-all-trades agents. Use dedicated Planner, Generator, Evaluator.",[186,2470,2471,2474],{},[636,2472,2473],{},"Validation loops",": External checks (tests, linters) before declaring \"done.\"",[186,2476,2477,2480],{},[636,2478,2479],{},"Externalized state",": Progress files in JSON or Markdown persist across resets.",[13,2482,2483,2484,2489,2490,2495],{},"Without it, 88% of agent projects never hit production—not from dumb models, but brittle systems.",[17,2485,2486],{},[20,2487,26],{"href":22,"ariaDescribedBy":2488,"dataFootnoteRef":8,"id":43},[24]," MindStudio calls this the shift from \"prompt hacking\" to \"session engineering.\"",[17,2491,2492],{},[20,2493,108],{"href":105,"ariaDescribedBy":2494,"dataFootnoteRef":8,"id":107},[24]," You need it for coding agents, long-running tasks, or anything beyond chat.",[13,2497,2498],{},"I've built agent pipelines for indie apps. Swapping Claude for Gemini barely moved the needle—until I added a harness. Reliability jumped from flaky to shippable.",[46,2500,2502],{"id":2501},"why-single-ai-agents-fail-in-production","Why Single AI Agents Fail in Production",[13,2504,2505,2506,2511],{},"Single AI agents fail production at an 88% rate because LLMs suffer generosity bias (self-praise of mediocre work), context anxiety (rushing as tokens fill), and unreliable self-evaluation—leading to broken outputs in coding or long tasks. Benchmarks show solo agents produce non-functional apps, while harnessed systems deliver working ones. Apurv Khare nails it: models can't judge their own code.",[17,2507,2508],{},[20,2509,34],{"href":31,"ariaDescribedBy":2510,"dataFootnoteRef":8,"id":127},[24]," (68 words)",[13,2513,2514,2515,2518],{},"Start with ",[636,2516,2517],{},"generosity bias",". LLMs rate their output high even when it's junk. A solo agent declares a half-baked function \"complete\" because it lacks external critique.",[13,2520,2521,2522,2525],{},"Then ",[636,2523,2524],{},"context anxiety",": As the window fills with history, models truncate reasoning to fit. Outputs get hasty, errors compound.",[13,2527,2528],{},"Self-evaluation? Worse. The same model generating code can't spot its flaws—it's like grading your own homework.",[13,2530,2531,2532,2537],{},"Real example: That 20-minute\u002F$9 solo agent game.",[17,2533,2534],{},[20,2535,34],{"href":31,"ariaDescribedBy":2536,"dataFootnoteRef":8,"id":162},[24]," It rendered sprites but crashed on collisions. No multiplayer, no saves—core broken.",[13,2539,2540],{},"Community echoes this. Nurunnubi Talukder: \"Generator-judge fails because models lie to themselves.\" Christian Ingul: \"Same agent generating and judging doesn't work.\" Divyansh Puri: \"A harness tames the wild horse.\"",[13,2542,2543],{},"For you? Wasted API spend, debugging agent hallucinations in prod, features that flake under load. Solo agents demo well but ship poorly.",[46,2545,2547],{"id":2546},"the-3-core-agent-harness-planner-generator-evaluator","The 3-Core-Agent Harness: Planner, Generator, Evaluator",[13,2549,2550,2551,2556],{},"The 3-core-agent harness in harness engineering separates concerns into a Planner (goal-to-spec), Generator (incremental builds with negotiated \"done\" contracts), and Evaluator (independent testing\u002Fcritique)—creating a GAN-inspired loop that overcomes solo agent limits for production reliability in tasks like game engines or apps. Anthropic's benchmarks prove it: harnessed agents ship quality where solos fail.",[17,2552,2553],{},[20,2554,34],{"href":31,"ariaDescribedBy":2555,"dataFootnoteRef":8,"id":236},[24]," (74 words)",[13,2558,2559],{},"Here's the flow:",[401,2561,2562,2567,2572,2577],{},[186,2563,2564,2566],{},[636,2565,1887],{},": High-level goal → detailed spec + Definition of Done (DoD).",[186,2568,2569,2571],{},[636,2570,1495],{},": Picks a sprint, negotiates DoD with Evaluator, codes incrementally using progress files.",[186,2573,2574,2576],{},[636,2575,1501],{},": Runs end-to-end tests (e.g., Playwright), reports file\u002Fline fixes needed.",[186,2578,2579],{},"Loop: Failures route back to Generator. Success advances progress file.",[13,2581,2582,2583,2588],{},"It's GAN-like: Generator improves via Evaluator's adversarial feedback.",[17,2584,2585],{},[20,2586,146],{"href":143,"ariaDescribedBy":2587,"dataFootnoteRef":8,"id":145},[24]," External memory—progress.json tracks tasks, code state, decisions—handles resets.",[13,2590,2591],{},"Why superior? Independent eval kills bias. Negotiation upfront aligns expectations. Iteration polishes.",[13,2593,2594,2595,2600],{},"From Anthropic: Solo = broken basics. Harness = advanced, playable.",[17,2596,2597],{},[20,2598,34],{"href":31,"ariaDescribedBy":2599,"dataFootnoteRef":8,"id":247},[24]," I've used variants for backend APIs—output went from 60% passing tests to 95%.",[13,2602,2603],{},"Prompt each core narrowly. Planner stays high-level to avoid over-spec. Generator reads progress.json. Evaluator gets tools for real testing.",[46,2605,2607],{"id":2606},"how-the-planner-turns-goals-into-actionable-specs","How the Planner Turns Goals into Actionable Specs",[13,2609,2610,2611,2616],{},"The Planner agent in harness engineering takes 1-4 sentence high-level goals and expands them into detailed product specs and Definition of Done, avoiding cascading errors by staying high-level and defining success criteria upfront—setting a solid foundation for Generator sprints. Anthropic's game engine example: \"Build a multiplayer shooter\" becomes UI flows, mechanics, tech stack, and testable DoD.",[17,2612,2613],{},[20,2614,34],{"href":31,"ariaDescribedBy":2615,"dataFootnoteRef":8,"id":354},[24]," (62 words)",[13,2618,2619],{},"Process is simple:",[183,2621,2622,2625],{},[186,2623,2624],{},"Input: \"Build a web-based task manager with user auth.\"",[186,2626,2627],{},"Output: YAML spec with features (login, CRUD tasks, search), stack (React\u002FNode\u002FPostgres), DoD (e.g., \"User can create task via UI, persists to DB, lists 100 items\").",[13,2629,2630],{},"Benefits? No Generator assumptions—everything explicit. Sprints stay bite-sized (one feature).",[13,2632,2633],{},"Example spec snippet:",[264,2635,2637],{"className":266,"code":2636,"language":268,"meta":8,"style":8},"features:\n  - login: JWT auth, form validation\ndod:\n  - \"Playwright test: New user logs in, redirects to dashboard\"\n  - \"API endpoint \u002Ftasks returns JSON array\"\n",[270,2638,2639,2646,2658,2665,2672],{"__ignoreMap":8},[273,2640,2641,2644],{"class":275,"line":276},[273,2642,2643],{"class":279},"features",[273,2645,284],{"class":283},[273,2647,2648,2650,2653,2655],{"class":275,"line":287},[273,2649,290],{"class":283},[273,2651,2652],{"class":279},"login",[273,2654,296],{"class":283},[273,2656,2657],{"class":299},"JWT auth, form validation\n",[273,2659,2660,2663],{"class":275,"line":303},[273,2661,2662],{"class":279},"dod",[273,2664,284],{"class":283},[273,2666,2667,2669],{"class":275,"line":314},[273,2668,290],{"class":283},[273,2670,2671],{"class":299},"\"Playwright test: New user logs in, redirects to dashboard\"\n",[273,2673,2674,2676],{"class":275,"line":724},[273,2675,290],{"class":283},[273,2677,2678],{"class":299},"\"API endpoint \u002Ftasks returns JSON array\"\n",[13,2680,2681],{},"Tips: Prompt with \"Output YAML only. Stay high-level—no code.\" Limit to 5-10 sprints. Review manually first time.",[13,2683,2684],{},"This prevents 80% of early failures I've seen. Builders skip it, watch cascades.",[46,2686,2688],{"id":2687},"generator-and-evaluator-the-iteration-loop-that-delivers-quality","Generator and Evaluator: The Iteration Loop That Delivers Quality",[13,2690,2691,2692,2697],{},"In harness engineering, the Generator implements sprints based on Planner specs and pre-negotiated DoD contracts, while the Evaluator independently tests via tools like Playwright—reporting precise file\u002Fline failures for targeted fixes in a feedback loop that yields far superior outputs than self-evaluation. Contrast: Solo agents break on basics; harnessed ones ship advanced features.",[17,2693,2694],{},[20,2695,34],{"href":31,"ariaDescribedBy":2696,"dataFootnoteRef":8,"id":1403},[24]," (71 words)",[13,2699,2700,2702],{},[636,2701,1495],{}," reads spec, progress.json, negotiates DoD:",[264,2704,2707],{"className":2705,"code":2706,"language":962},[960],"Generator: \"Sprint: Login form. Proposed DoD: Form submits, stores JWT in localStorage.\"\nEvaluator: \"Accept. Add: Error on invalid creds.\"\n",[270,2708,2706],{"__ignoreMap":8},[13,2710,2711],{},"Codes incrementally—edits files, commits to git-like state.",[13,2713,2714,2716],{},[636,2715,1501],{}," launches app, runs tests:",[183,2718,2719,2722,2725],{},[186,2720,2721],{},"UI: Playwright clicks, asserts.",[186,2723,2724],{},"API: Curl endpoints.",[186,2726,2727],{},"DB: Query state.",[13,2729,2730],{},"Feedback: \"src\u002Fauth.js:47—JWT decode fails on expiry. Fix expiry check.\"",[13,2732,2733],{},"Loop 3-5 times per sprint. Tools make it concrete—no vagueness.",[13,2735,2736,2737],{},"GAN analogy from Talukder: Evaluator \"fights\" Generator to quality.",[17,2738,2739],{},[20,2740,146],{"href":143,"ariaDescribedBy":2741,"dataFootnoteRef":8,"id":155},[24],[13,2743,2744,2745,2750],{},"Solo vs. harness: 20min broken → 6hr playable.",[17,2746,2747],{},[20,2748,34],{"href":31,"ariaDescribedBy":2749,"dataFootnoteRef":8,"id":1452},[24]," In my tests, eval precision cut revisions 40%.",[13,2752,2753],{},"Integrate Playwright:",[264,2755,2757],{"className":1011,"code":2756,"language":1013,"meta":8,"style":8},"npx playwright test --project=agent-eval\n",[270,2758,2759],{"__ignoreMap":8},[273,2760,2761,2764,2767,2770],{"class":275,"line":276},[273,2762,2763],{"class":1020},"npx",[273,2765,2766],{"class":299}," playwright",[273,2768,2769],{"class":299}," test",[273,2771,2772],{"class":1024}," --project=agent-eval\n",[46,2774,2776],{"id":2775},"harness-components-guides-sensors-and-orchestration","Harness Components: Guides, Sensors, and Orchestration",[13,2778,2779,2780,2785],{},"Effective harness engineering relies on guides (feedforward docs\u002Fspecs), sensors (feedback like linters\u002Ftests), session orchestration (context injection\u002Fvalidation), and output loops—plus state files and specialization—to make agents reliable across resets and complex tasks. Tian Pan: Guides steer proactively; sensors correct reactively.",[17,2781,2782],{},[20,2783,79],{"href":76,"ariaDescribedBy":2784,"dataFootnoteRef":8,"id":181},[24]," (54 words)",[13,2787,2788,2791],{},[636,2789,2790],{},"Guides",": Docs encoding \"good\"—architecture.md, prompt templates, examples. Injected per session.",[13,2793,2794,2797],{},[636,2795,2796],{},"Sensors",": Linters (ESLint), tests (Jest), AI reviewers. Run post-generation.",[13,2799,2800,2803,2804],{},[636,2801,2802],{},"Orchestration"," (MindStudio):",[17,2805,2806],{},[20,2807,108],{"href":105,"ariaDescribedBy":2808,"dataFootnoteRef":8,"id":117},[24],[183,2810,2811,2814,2817,2820],{},[186,2812,2813],{},"Start: Inject goal + progress.",[186,2815,2816],{},"Route: Planner → Generator\u002FEval loop.",[186,2818,2819],{},"Validate: Schema check outputs.",[186,2821,2822],{},"Persist: Update progress.json.",[13,2824,2825],{},"State files example:",[264,2827,2829],{"className":1182,"code":2828,"language":1184,"meta":8,"style":8},"{\n  \"completed\": [\"auth\"],\n  \"in_progress\": \"tasks-crud\",\n  \"dod\": [\"API returns 200 with tasks\"]\n}\n",[270,2830,2831,2835,2847,2859,2871],{"__ignoreMap":8},[273,2832,2833],{"class":275,"line":276},[273,2834,1191],{"class":283},[273,2836,2837,2840,2842,2845],{"class":275,"line":287},[273,2838,2839],{"class":1024},"  \"completed\"",[273,2841,730],{"class":283},[273,2843,2844],{"class":299},"\"auth\"",[273,2846,1213],{"class":283},[273,2848,2849,2852,2854,2857],{"class":275,"line":303},[273,2850,2851],{"class":1024},"  \"in_progress\"",[273,2853,296],{"class":283},[273,2855,2856],{"class":299},"\"tasks-crud\"",[273,2858,1226],{"class":283},[273,2860,2861,2864,2866,2869],{"class":275,"line":314},[273,2862,2863],{"class":1024},"  \"dod\"",[273,2865,730],{"class":283},[273,2867,2868],{"class":299},"\"API returns 200 with tasks\"",[273,2870,736],{"class":283},[273,2872,2873],{"class":275,"line":724},[273,2874,1245],{"class":283},[13,2876,2877],{},"Specialization: Add ArchAgent for design, TestGen for suites.",[13,2879,2880],{},"This combo handles resets—your session crashed? Resume from state.",[13,2882,2883],{},"No guides\u002Fsensors? Back to solo fragility.",[46,2885,2887],{"id":2886},"real-results-why-the-harness-beats-solo-agents","Real Results: Why the Harness Beats Solo Agents",[13,2889,2890,2891,2896],{},"Benchmarks from Anthropic show the 3-core harness builds fully playable games with advanced features in 6 hours\u002F$200, versus solo agents' broken cores in 20min\u002F$9—proving harness engineering delivers production quality where models alone fail. Community validates: Better systems, not smarter models.",[17,2892,2893],{},[20,2894,34],{"href":31,"ariaDescribedBy":2895,"dataFootnoteRef":8,"id":1562},[24]," (52 words)",[13,2898,2899],{},"Anthropic case: Goal—\"multiplayer FPS.\"",[183,2901,2902,2905],{},[186,2903,2904],{},"Solo: Sprites move, but no collisions, scoring, multiplayer.",[186,2906,2907],{},"Harness: Full UI, netcode, AI bots, saves—all working.",[13,2909,2910,2911,2916,2917,2923],{},"Khare: \"Harness is the interesting space.\"",[17,2912,2913],{},[20,2914,34],{"href":31,"ariaDescribedBy":2915,"dataFootnoteRef":8,"id":1571},[24]," Talukder: \"Encode model limits.\"",[17,2918,2919],{},[20,2920,146],{"href":143,"ariaDescribedBy":2921,"dataFootnoteRef":8,"id":2922},[24],"user-content-fnref-5-3"," Ripley: \"Systems around models.\" Ben Ripley on LinkedIn.",[13,2925,2926,2927],{},"Broader: Fixes 88% failure by design.",[17,2928,2929],{},[20,2930,26],{"href":22,"ariaDescribedBy":2931,"dataFootnoteRef":8,"id":58},[24],[13,2933,2934],{},"Evolve it: As models improve, delete sensors for self-eval. Audit quarterly.",[13,2936,2937],{},"I've shipped a CRM agent this way—solo would've bombed prod.",[46,2939,2941],{"id":2940},"trade-offs-and-how-to-implement-harness-engineering-today","Trade-offs and How to Implement Harness Engineering Today",[13,2943,2944,2945,2950],{},"Harness engineering adds cost\u002Flatency (e.g., 30x time) but justifies for production via reliability; implement via LangChain\u002FCrewAI for orchestration, start with Generator-Evaluator pair for simple tasks, scale to 3-core, and audit regularly. Not frameworks alone—custom loops win.",[17,2946,2947],{},[20,2948,79],{"href":76,"ariaDescribedBy":2949,"dataFootnoteRef":8,"id":194},[24]," (48 words)",[13,2952,1536],{},[183,2954,2955,2958,2961],{},[186,2956,2957],{},"Overhead: $200 vs. $9, hours vs. minutes.",[186,2959,2960],{},"When simplify: Short tasks → Gen-Eval. Complex → full harness.",[186,2962,2963],{},"Cost threshold: >$50\u002Ftask? Harness pays off.",[13,2965,2966],{},"Misconception: CrewAI ≠ harness. It's orchestration—build custom.",[13,2968,2969],{},"Quick start pseudocode:",[264,2971,2973],{"className":1266,"code":2972,"language":1268,"meta":8,"style":8},"while not done(progress):\n    spec = planner(goal)\n    for sprint in spec:\n        dod = negotiate(gen, eval, sprint)\n        code = generator(sprint, progress)\n        feedback = evaluator(code, dod)\n        if pass: update_progress()\n",[270,2974,2975,2980,2985,2990,2995,3000,3005],{"__ignoreMap":8},[273,2976,2977],{"class":275,"line":276},[273,2978,2979],{},"while not done(progress):\n",[273,2981,2982],{"class":275,"line":287},[273,2983,2984],{},"    spec = planner(goal)\n",[273,2986,2987],{"class":275,"line":303},[273,2988,2989],{},"    for sprint in spec:\n",[273,2991,2992],{"class":275,"line":314},[273,2993,2994],{},"        dod = negotiate(gen, eval, sprint)\n",[273,2996,2997],{"class":275,"line":724},[273,2998,2999],{},"        code = generator(sprint, progress)\n",[273,3001,3002],{"class":275,"line":739},[273,3003,3004],{},"        feedback = evaluator(code, dod)\n",[273,3006,3007],{"class":275,"line":752},[273,3008,3009],{},"        if pass: update_progress()\n",[13,3011,3012],{},"Tools: LangChain for chains, Playwright for eval, JSON for state.",[13,3014,3015],{},"Audit: Run solo vs. harness on toy task. Measure pass rate.",[46,3017,3019],{"id":3018},"next-steps-bootstrap-your-harness-today","Next Steps: Bootstrap Your Harness Today",[13,3021,3022],{},"Harness engineering moves agents from demo toys to production engines—Planner sets rails, Generator-Eval loop grinds quality, components make it resilient. Recap: Model matters less than system.",[13,3024,3025],{},"Future: Dynamic harnesses auto-simplify; agent-first dev.",[13,3027,3028],{},"Checklist:",[183,3030,3031,3034,3037,3040,3043],{},[186,3032,3033],{},"Map your agent to 3 cores.",[186,3035,3036],{},"Add Playwright to Eval.",[186,3038,3039],{},"Externalize state to JSON.",[186,3041,3042],{},"Negotiate DoD pre-code.",[186,3044,3045],{},"Benchmark solo vs. harness on a feature.",[13,3047,3048],{},"Audit your current project now: Slot it into Planner\u002FGen\u002FEval, wire an independent evaluator with Playwright tests, run one sprint, and compare output quality. You'll see the gap—and close it.",[391,3050,3052,3055],{"className":3051,"dataFootnotes":8},[394],[46,3053,399],{"className":3054,"id":24},[398],[401,3056,3057,3078,3133,3153,3168],{},[186,3058,3059,414,3065,414,3068,414,3073],{"id":405},[20,3060,3064],{"href":3061,"rel":3062},"https:\u002F\u002Fmiraflow.ai\u002Fblog\u002Fharness-engineering-why-88-percent-ai-agents-fail",[3063],"nofollow","Miraflow AI: Harness Engineering: Why 88% of AI Agents Fail",[20,3066,413],{"href":409,"ariaLabel":410,"className":3067,"dataFootnoteBackref":8},[412],[20,3069,413,3071],{"href":417,"ariaLabel":418,"className":3070,"dataFootnoteBackref":8},[412],[17,3072,34],{},[20,3074,413,3076],{"href":424,"ariaLabel":425,"className":3075,"dataFootnoteBackref":8},[412],[17,3077,79],{},[186,3079,3080,414,3085,414,3088,414,3093,414,3098,414,3103,414,3108,414,3113,414,3118,414,3123,414,3128],{"id":460},[20,3081,3084],{"href":3082,"rel":3083},"https:\u002F\u002Fwww.linkedin.com\u002Fposts\u002Fapurvkhare_apurv-khare-harness-engineering-long-activity-7446406273853562880-khax",[3063],"Apurv Khare on LinkedIn: Harness Engineering",[20,3086,413],{"href":464,"ariaLabel":465,"className":3087,"dataFootnoteBackref":8},[412],[20,3089,413,3091],{"href":469,"ariaLabel":470,"className":3090,"dataFootnoteBackref":8},[412],[17,3092,34],{},[20,3094,413,3096],{"href":476,"ariaLabel":477,"className":3095,"dataFootnoteBackref":8},[412],[17,3097,79],{},[20,3099,413,3101],{"href":483,"ariaLabel":484,"className":3100,"dataFootnoteBackref":8},[412],[17,3102,108],{},[20,3104,413,3106],{"href":490,"ariaLabel":491,"className":3105,"dataFootnoteBackref":8},[412],[17,3107,146],{},[20,3109,413,3111],{"href":497,"ariaLabel":498,"className":3110,"dataFootnoteBackref":8},[412],[17,3112,343],{},[20,3114,413,3116],{"href":1667,"ariaLabel":1668,"className":3115,"dataFootnoteBackref":8},[412],[17,3117,457],{},[20,3119,413,3121],{"href":1674,"ariaLabel":1675,"className":3120,"dataFootnoteBackref":8},[412],[17,3122,1633],{},[20,3124,413,3126],{"href":1681,"ariaLabel":1682,"className":3125,"dataFootnoteBackref":8},[412],[17,3127,1686],{},[20,3129,413,3131],{"href":1689,"ariaLabel":1690,"className":3130,"dataFootnoteBackref":8},[412],[17,3132,1694],{},[186,3134,3135,414,3140,414,3143,414,3148],{"id":504},[20,3136,3139],{"href":3137,"rel":3138},"https:\u002F\u002Ftianpan.co\u002Fblog\u002F2026-02-17-harness-engineering-agent-first-software-development",[3063],"Tian Pan: Harness Engineering",[20,3141,413],{"href":508,"ariaLabel":509,"className":3142,"dataFootnoteBackref":8},[412],[20,3144,413,3146],{"href":513,"ariaLabel":514,"className":3145,"dataFootnoteBackref":8},[412],[17,3147,34],{},[20,3149,413,3151],{"href":520,"ariaLabel":521,"className":3150,"dataFootnoteBackref":8},[412],[17,3152,79],{},[186,3154,3155,414,3160,414,3163],{"id":534},[20,3156,3159],{"href":3157,"rel":3158},"https:\u002F\u002Fwww.mindstudio.ai\u002Fblog\u002Fwhat-is-harness-engineering-ai-coding",[3063],"MindStudio: What Is Harness Engineering?",[20,3161,413],{"href":538,"ariaLabel":539,"className":3162,"dataFootnoteBackref":8},[412],[20,3164,413,3166],{"href":543,"ariaLabel":544,"className":3165,"dataFootnoteBackref":8},[412],[17,3167,34],{},[186,3169,3170,414,3175,414,3178,414,3183],{"id":550},[20,3171,3174],{"href":3172,"rel":3173},"https:\u002F\u002Fwww.linkedin.com\u002Fposts\u002Fnurunnubi_harness-design-for-long-running-application-activity-7447849188882632704--iEm",[3063],"Nurunnubi Talukder on LinkedIn",[20,3176,413],{"href":554,"ariaLabel":555,"className":3177,"dataFootnoteBackref":8},[412],[20,3179,413,3181],{"href":559,"ariaLabel":560,"className":3180,"dataFootnoteBackref":8},[412],[17,3182,34],{},[20,3184,413,3188],{"href":3185,"ariaLabel":3186,"className":3187,"dataFootnoteBackref":8},"#user-content-fnref-5-3","Back to reference 5-3",[412],[17,3189,79],{},[588,3191,3192],{},"html pre.shiki code .s9eBZ, html code.shiki .s9eBZ{--shiki-default:#22863A;--shiki-dark:#85E89D}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}",{"title":8,"searchDepth":287,"depth":287,"links":3194},[3195,3196,3197,3198,3199,3200,3201,3202,3203,3204],{"id":2441,"depth":287,"text":2442},{"id":2501,"depth":287,"text":2502},{"id":2546,"depth":287,"text":2547},{"id":2606,"depth":287,"text":2607},{"id":2687,"depth":287,"text":2688},{"id":2775,"depth":287,"text":2776},{"id":2886,"depth":287,"text":2887},{"id":2940,"depth":287,"text":2941},{"id":3018,"depth":287,"text":3019},{"id":24,"depth":287,"text":399},"Production AI agents fail at an 88% rate despite LLM advances because they rely on solo models without proper scaffolding.1 Harness engineering fixes this: build a 3-core system with a Planner for specs, Generator for code, and independent Evaluator for critique. This outperforms single agents by countering generosity bias and self-evaluation flaws (42 words).",[607,3207,3208,1782,3209,3210,3211,3212],"AI agents","production AI","AI agent failure","LLM harness","agent architecture","reliable AI systems",{},"Discover harness engineering: Build reliable AI agents with Planner, Generator, and Evaluator to overcome 88% production failure rates. Proven in Anthropic's game engine benchmarks for coding tasks.","Harness Engineering for Production AI Agents","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-baseline-round2-article",[],{"intro":1793,"what-is-harness-engineering":2442,"why-single-ai-agents-fail-in-production":2502,"the-3-core-agent-harness-planner-generator-evaluator":2547,"how-the-planner-turns-goals-into-actionable-specs":2607,"generator-and-evaluator-the-iteration-loop-that-delivers-quality":2688,"harness-components-guides-sensors-and-orchestration":2776,"real-results-why-the-harness-beats-solo-agents":2887,"trade-offs-and-how-to-implement-harness-engineering-today":2941},{"title":2414,"description":3205},{"loc":3216},"harness-engineering-3-core-system-reliable-production-ai-agents","articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-baseline-round2-article","9F3Tc1PUu6u5THLqevKsmyTtniK0N-jFEnyxvezgOFU",{"id":3225,"title":3226,"ai_summary":3227,"author":8,"body":3228,"category":3369,"date_modified":601,"description":3232,"excerpt":8,"extension":603,"faq":601,"keywords":3370,"meta":3378,"meta_description":3379,"meta_title":3380,"navigation":616,"path":3381,"primary_keyword":3371,"published_at":601,"related_posts":3382,"sections":3383,"seo":3384,"sitemap":3385,"slug":3386,"status":623,"stem":3387,"__hash__":3388},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expa-balanced-post-merge-article.md","The 3-Core-Agent Harness: Why Production AI Needs Governance, Not Just Frameworks","The 3-core-agent harness is an architectural pattern that splits AI tasks into specialized roles: a Planner (specifications), a Generator (execution), and an Evaluator (testing). This separation prevents self-evaluation bias and ensures production-grade reliability by mirroring the software engineering sprint cycle of scope, execute, and QA to deliver hardened results.",{"type":10,"value":3229,"toc":3361},[3230,3233,3236,3240,3243,3246,3266,3269,3273,3276,3279,3290,3294,3297,3300,3303,3307,3310,3313,3316,3330,3334,3337,3340,3343,3351,3355,3358],[13,3231,3232],{},"The 3-core-agent harness is an architectural pattern that splits AI tasks into specialized roles: a Planner (specifications), a Generator (execution), and an Evaluator (testing). This separation prevents self-evaluation bias and ensures production-grade reliability by mirroring the software engineering sprint cycle of scope, execute, and QA.",[13,3234,3235],{},"In early 2024, an autonomous AI agent successfully hacked a FreeBSD system in just four hours. It didn't do this because it was \"evil\" or sentient; it did it because it was given a goal and zero governance to check its work. When we build monolithic agents that plan, execute, and grade their own homework, we aren't building assets—we are building liabilities. The shift from \"cool demos\" to production-grade AI requires moving away from the black-box agent and toward a structured, role-based harness.",[46,3237,3239],{"id":3238},"the-3-core-agent-harness-a-blueprint-for-production-reliability","The 3-Core-Agent Harness: A Blueprint for Production Reliability",[13,3241,3242],{},"The 3-core-agent harness replaces the \"black box\" agent approach with a transparent, three-tiered hierarchy of Planner, Generator, and Evaluator. By isolating the specification from execution and testing, builders can force agents to negotiate strict technical contracts that define success before work begins. This structured separation ensures the system meets high-level requirements and satisfies rigorous acceptance criteria, preventing the cascading hallucinations common in monolithic models.",[13,3244,3245],{},"This architecture treats the AI pipeline as a professional engineering team rather than a single, overworked intern. The roles are strictly defined:",[183,3247,3248,3254,3260],{},[186,3249,3250,3253],{},[636,3251,3252],{},"The Planner",": This agent takes a vague, two-sentence prompt and converts it into a hardened technical specification. It prevents cascading assumptions by defining the boundaries of the task before work begins.",[186,3255,3256,3259],{},[636,3257,3258],{},"The Generator",": This is the \"worker\" that executes the sprint. It builds against the Planner’s contract using modern stacks like Vite, FastAPI, or React, focusing purely on implementation without the distraction of high-level strategy.",[186,3261,3262,3265],{},[636,3263,3264],{},"The Evaluator",": The \"skeptic\" of the group. It doesn't just look at the code; it uses concrete tools like Playwright to verify output against the original spec, providing line-level failure reports.",[13,3267,3268],{},"By using this harness, you ensure that the system remains observable. When a failure occurs, you can pinpoint exactly which role dropped the ball: the plan was flawed, the execution missed the mark, or the evaluation was too lenient.",[46,3270,3272],{"id":3271},"why-the-evaluator-is-your-most-critical-hire","Why the Evaluator is Your Most Critical Hire",[13,3274,3275],{},"Anthropic’s research on multi-agent systems reveals that agents suffer from \"generosity bias\" and \"context anxiety,\" causing them to overrate their own mediocre work as their context windows fill up. A dedicated Evaluator agent solves this by operating as an independent QA layer, providing the granular, line-level feedback necessary for meaningful iteration.",[13,3277,3278],{},"Generosity bias is a documented phenomenon where a model ignores its own errors to maintain the appearance of progress. When the same model that wrote the code is asked to review it, it consistently misses bugs it introduced itself. Furthermore, \"context anxiety\" causes models to wrap up tasks prematurely as they approach their token limits, leading to \"good enough\" solutions that break in production.",[13,3280,3281,3282,3285,3286,3289],{},"A separate Evaluator agent changes the feedback loop from vague commentary to contract-based failure reports. Instead of saying \"the UI looks a bit off,\" a dedicated Evaluator using a tool like Playwright can report: \"",[270,3283,3284],{},"fillRectangle"," exists in the source but isn't triggered on ",[270,3287,3288],{},"mouseUp",".\" This level of specificity allows the Generator to perform a surgical fix rather than guessing at the problem.",[46,3291,3293],{"id":3292},"harnesses-vs-frameworks-trading-control-for-reliable-output","Harnesses vs. Frameworks: Trading Control for Reliable Output",[13,3295,3296],{},"While agent frameworks like CrewAI or LangChain provide flexible building blocks, a 3-core-agent harness is a maximally opinionated system designed for specific, reliable outcomes. Frameworks are for builders who want to design the engine from scratch; harnesses provide users with a complete vehicle featuring pre-configured navigation, braking systems, and automated safety checks.",[13,3298,3299],{},"The distinction lies in the \"batteries-included\" nature of a harness. In a framework, you must decide how agents communicate, how memory is handled, and how loops are structured. In a harness like OpenClaw, these decisions are pre-configured. You are trading the ability to customize every internal gear for the assurance that the system will not exceed its mandate or hallucinate a success state.",[13,3301,3302],{},"For most product-minded builders, a harness is the superior choice for production. It provides a stable environment where memory, context, and agent interactions are already optimized for high-complexity tasks, allowing you to focus on the product features rather than the orchestration logic.",[46,3304,3306],{"id":3305},"the-9-vs-200-decision-managing-the-cost-quality-tradeoff","The $9 vs. $200 Decision: Managing the Cost-Quality Tradeoff",[13,3308,3309],{},"Production-grade AI is not cheap; a solo agent run might cost $9 and take 20 minutes, but a full 3-core-agent harness run can exceed $200 over six hours. This 40x cost difference, highlighted in analysis by engineering leader Apurv Khare, forces builders to treat agent orchestration like a PM’s sprint planning, where expensive resources are reserved specifically for high-complexity features.",[13,3311,3312],{},"Khare's data demonstrates that while the cost jump is significant, the delta in quality is massive. A solo run often results in \"broken entity wiring\" where components exist but don't talk to each other. A full harness run, involving 5 to 15 iterations, can produce a 16-feature working application with sound effects, sprite generation, and shareable exports.",[13,3314,3315],{},"The decision to use a harness should follow a simple complexity threshold:",[183,3317,3318,3324],{},[186,3319,3320,3323],{},[636,3321,3322],{},"Low Complexity (\u003C$10):"," Simple scripts, unit tests, or content drafts. Use a solo agent.",[186,3325,3326,3329],{},[636,3327,3328],{},"High Complexity (>$150):"," Full feature builds, complex UI interactions, or multi-step data pipelines. Use the 3-core-agent harness.",[46,3331,3333],{"id":3332},"implementing-governance-from-simple-prompts-to-sprint-cycles","Implementing Governance: From Simple Prompts to Sprint Cycles",[13,3335,3336],{},"Implementing a 3-core-agent harness effectively turns your AI pipeline into a 5-to-15 iteration sprint cycle. Each loop refines the output as the Evaluator identifies failures and the Planner adjusts the spec.",[13,3338,3339],{},"This iterative process allows the Generator to attempt fresh implementations based on precise feedback. It effectively mirrors the traditional software development lifecycle but executes at machine speed.",[13,3341,3342],{},"This cycle is the only way to prevent the \"FreeBSD scenario,\" where an ungoverned agent autonomously exploited a system in four hours. By baking scope constraints and permission boundaries into the Planner role, you ensure the autonomous system cannot wander outside its designated sandbox. The Evaluator then acts as the final gatekeeper, refusing to \"merge\" the work until every acceptance criterion is met.",[13,3344,3345,3346,3350],{},"Success in this model depends on negotiating \"Done\" contracts ",[3347,3348,3349],"em",{},"before"," generation starts. If the Planner and Evaluator agree on what success looks like—down to the specific API endpoints and UI states—the Generator is much more likely to hit the target on the first or second iteration.",[46,3352,3354],{"id":3353},"conclusion","Conclusion",[13,3356,3357],{},"The 3-core-agent harness is the recognition that AI agents require the same governance and separation of concerns as human engineering teams. By decoupling planning, execution, and skepticism, we move from \"cool demos\" to reliable, autonomous production systems.",[13,3359,3360],{},"Audit your current agentic workflows: identify where a single agent is currently \"grading its own homework\" and experiment with a separate, skeptically-prompted Evaluator agent to review its next five outputs.",{"title":8,"searchDepth":287,"depth":287,"links":3362},[3363,3364,3365,3366,3367,3368],{"id":3238,"depth":287,"text":3239},{"id":3271,"depth":287,"text":3272},{"id":3292,"depth":287,"text":3293},{"id":3305,"depth":287,"text":3306},{"id":3332,"depth":287,"text":3333},{"id":3353,"depth":287,"text":3354},"AI Engineering",[3371,3372,3373,3374,2517,3375,3376,3377],"3-core-agent harness","AI agent governance","agentic workflows","Anthropic research","autonomous systems","AI software engineering","primary_keyword: 3-core-agent harness",{},"Discover why splitting AI agents into Planner, Generator, and Evaluator roles is the only way to eliminate bias and ship production-grade autonomous systems.","The 3-Core-Agent Harness for Production AI Reliability","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expa-balanced-post-merge-article",[],{"intro":2407,"blueprint":3239,"evaluator":3272,"harness-vs-framework":3293,"cost-tradeoff":3306,"governance":3333,"conclusion":3354},{"title":3226,"description":3232},{"loc":3381},"3-core-agent-harness-production-reliability","articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expa-balanced-post-merge-article","uJNkAB5y-PSAv8pt1_N-vNENos2JrZho0VX37EYihGQ",{"id":3390,"title":3391,"ai_summary":3392,"author":8,"body":3393,"category":3369,"date_modified":601,"description":3524,"excerpt":8,"extension":603,"faq":601,"keywords":3525,"meta":3531,"meta_description":3532,"meta_title":3533,"navigation":616,"path":3534,"primary_keyword":3371,"published_at":601,"related_posts":3535,"sections":3536,"seo":3537,"sitemap":3538,"slug":3539,"status":623,"stem":3540,"__hash__":3541},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expa-round2-article.md","The 3-Core-Agent Harness: Why Production Agents Need Structure, Not Just Models","The 3-core-agent harness is a production architecture that splits AI tasks into three specialized roles: Planner, Generator, and Evaluator. This separation prevents context drift and self-evaluation bias, allowing systems to complete complex, multi-hour tasks that single-agent systems fail to execute reliably due to lack of internal skepticism.",{"type":10,"value":3394,"toc":3516},[3395,3401,3404,3408,3414,3417,3420,3424,3427,3433,3441,3445,3448,3451,3477,3480,3484,3487,3490,3493,3497,3500,3503,3505,3508,3511],[13,3396,3397,3398,3400],{},"Single-agent systems are great for demos but reliably fail in production because they are essentially developers forced to be their own product managers and QA engineers simultaneously. The ",[636,3399,3371],{}," is a production-grade architecture that separates concerns into specialized Planner, Generator, and Evaluator agents. This structural division prevents context loss and self-evaluation bias, enabling AI systems to complete complex, multi-hour tasks that single agents fail to execute reliably.",[13,3402,3403],{},"We have all seen the \"cool demo\" where an agent builds a landing page in sixty seconds. However, when you ask that same agent to manage a four-hour migration or build a full-stack feature, the wheels come off. The failure isn't usually the model's intelligence; it is the lack of architectural constraints. Without a harness, agents suffer from context window saturation and a \"rubber-stamping\" effect where they stop being critical of their own errors.",[46,3405,3407],{"id":3406},"solving-the-context-coherence-and-self-evaluation-bias","Solving the Context Coherence and Self-Evaluation Bias",[13,3409,3410,3411,3413],{},"Single-agent failures stem from two psychological-parallel flaws: context coherence loss, where the model forgets its original goal during long sessions, and self-evaluation bias, where models consistently overpraise their own mediocre outputs. A ",[636,3412,3371],{}," solves this by externalizing the \"goal\" and \"critique\" into separate model instances to ensure the agent doing the work is never checking its own work.",[13,3415,3416],{},"In sessions lasting four hours or more, \"context drift\" becomes an inevitability. As the context window fills with logs, intermediate code snippets, and conversational filler, the original requirement loses its \"gravity\" in the model’s attention mechanism. Anthropic’s 2026 research, \"Harness Design for Long-Running Application Development,\" indicates that agents tend to respond with confident praise when asked to evaluate their own work, even when the quality is objectively poor.",[13,3418,3419],{},"Simply increasing the context window does not fix this. Reasoning decay happens because the model is trying to juggle too many conflicting internal roles. By splitting the work, you ensure the Planner stays focused on the \"North Star,\" while the Generator focuses on the immediate technical implementation.",[46,3421,3423],{"id":3422},"the-anatomy-of-a-3-core-agent-harness","The Anatomy of a 3-Core-Agent Harness",[13,3425,3426],{},"A robust 3-core-agent harness functions like a professional software development team, with the Planner defining the \"what,\" the Generator executing the \"how,\" and the Evaluator verifying those results against the original specification. This intentional separation ensures that the technical implementation never drifts from the business requirements by mirroring the classic software development lifecycle at machine speed.",[13,3428,3429,3430,3432],{},"The ",[636,3431,1887],{}," agent acts as the product manager. It converts brief, often vague prompts into high-level specifications. The goal here is clarity without over-specification; if the Planner dictates every line of code, it creates cascading errors. Instead, it defines what \"correct\" looks like, often aiming for a higher qualitative bar to push the Generator further.",[13,3434,3429,3435,3437,3438,3440],{},[636,3436,1495],{}," is the \"developer\" in the box. It works in focused sprints, using standard tools like Git, Vite, or a CLI rather than just outputting raw text. Finally, the ",[636,3439,1501],{}," must be architecturally separate to maintain skepticism. It uses runtime validation—like Playwright for UI tests or API calls—to verify that the Generator actually met the Planner's requirements.",[46,3442,3444],{"id":3443},"the-gan-inspired-feedback-loop-turning-rubrics-into-loss-functions","The GAN-Inspired Feedback Loop: Turning Rubrics into Loss Functions",[13,3446,3447],{},"The system operates as a generative adversarial network (GAN) where the Evaluator’s skeptical assessment drives the Generator's iterative improvement. By using weighted rubrics as \"loss functions,\" the harness transforms subjective quality judgments into objective, optimizable scoring signals. This creates a closed-loop system that refuses to accept \"good enough\" while maintaining a high bar for production-ready code.",[13,3449,3450],{},"The cycle follows a strict path:",[183,3452,3453,3459,3465,3471],{},[186,3454,3455,3458],{},[636,3456,3457],{},"Spec",": The Planner sets the target.",[186,3460,3461,3464],{},[636,3462,3463],{},"Implementation",": The Generator builds a version.",[186,3466,3467,3470],{},[636,3468,3469],{},"Skeptical Review",": The Evaluator attempts to break it.",[186,3472,3473,3476],{},[636,3474,3475],{},"Refinement",": The Generator receives structured feedback and tries again.",[13,3478,3479],{},"This loop relies on \"sprint contracts\"—shared definitions of \"done\" between agents. Rather than giving vague feedback like \"make the UI better,\" the Evaluator provides structured PASS\u002FFAIL criteria based on the rubric. If the Evaluator finds that a button doesn't trigger a modal, the Generator receives a specific failure report rather than a generic prompt to \"fix the bugs.\"",[46,3481,3483],{"id":3482},"production-patterns-from-stripes-minions-to-virtualized-sandboxes","Production Patterns: From Stripe’s Minions to Virtualized Sandboxes",[13,3485,3486],{},"Leading engineering teams at Stripe and OpenAI have moved beyond generic frameworks to custom harnesses that treat agents as stateless \"cattle\" rather than fragile \"pets.\" These systems use centralized toolservers and isolated devboxes to ensure every agent action is observable, reproducible, and verifiable across complex, multi-step engineering workflows.",[13,3488,3489],{},"Stripe’s \"Minions\" system is a prime example, shipping over 1,300 pull requests per week as detailed in Jasnova's 2026 analysis of harness engineering. They use a centralized Model Context Protocol (MCP) server called \"Toolshed\" to give agents access to over 500 internal tools. Each agent operates in a virtualized \"devbox\" that spins up in seconds, ensuring that if an agent messes up the environment, the harness can simply reboot it.",[13,3491,3492],{},"Similarly, OpenAI has integrated browser validation into agent runtimes. By using the Chrome DevTools Protocol, the Evaluator role can take screenshots, inspect the DOM, and verify that a fix actually works in a real browser. This level of observability is what separates a toy from a tool.",[46,3494,3496],{"id":3495},"the-20x-cost-premium-balancing-the-3-core-agent-harness-economics","The 20x Cost Premium: Balancing the 3-Core-Agent Harness Economics",[13,3498,3499],{},"Implementing a 3-core-agent harness involves a significant cost trade-off, often increasing token usage by 20x compared to solo agents. However, this investment represents the difference between a $9 broken prototype and a $200 fully functional production application. This cost premium facilitates the transition from unusable code to a shipped, feature-complete product that requires zero human intervention.",[13,3501,3502],{},"The baseline for multi-agent systems is roughly 3x the token burn of a single agent, but the iterative loops often push this much higher. As foundation models improve, engineers can follow the \"Shrinking Harness\" principle—removing components that the model can now handle natively. However, for high-stakes production tasks, the cost of a failed execution is almost always higher than the cost of the extra tokens.",[46,3504,3354],{"id":3353},[13,3506,3507],{},"The 3-core-agent harness proves that the \"intelligence\" of an AI system is as much a product of its architectural constraints as it is the underlying model. By separating planning, generation, and evaluation, builders can finally ship agents that handle long-running, complex tasks without degrading. This shift from monolithic prompts to structured workflows is the key to moving beyond brittle demos.",[13,3509,3510],{},"Audit your current agent implementation for \"self-grading.\" If your agent is evaluating its own output, try splitting that evaluation into a separate, skeptical model instance using a structured rubric and observe the immediate delta in quality. Separating the \"doer\" from the \"checker\" is the fastest way to move an AI project from a demo to a shipped product.",[13,3512,3513],{},[636,3514,3515],{},"Ready to build more reliable agents? Start by decoupling your QA logic from your generation prompt and implementing a dedicated Evaluator role in your next agentic workflow.",{"title":8,"searchDepth":287,"depth":287,"links":3517},[3518,3519,3520,3521,3522,3523],{"id":3406,"depth":287,"text":3407},{"id":3422,"depth":287,"text":3423},{"id":3443,"depth":287,"text":3444},{"id":3482,"depth":287,"text":3483},{"id":3495,"depth":287,"text":3496},{"id":3353,"depth":287,"text":3354},"Single-agent systems are great for demos but reliably fail in production because they are essentially developers forced to be their own product managers and QA engineers simultaneously. The 3-core-agent harness is a production-grade architecture that separates concerns into specialized Planner, Generator, and Evaluator agents. This structural division prevents context loss and self-evaluation bias, enabling AI systems to complete complex, multi-hour tasks that single agents fail to execute reliably.",[3371,3526,3373,3527,3528,3529,3530],"AI agent architecture","LLM evaluation bias","context coherence","Anthropic agent research","AI engineering patterns",{},"Learn why the 3-core-agent harness (Planner, Generator, Evaluator) is the standard for production AI systems to solve context drift and self-evaluation bias.","The 3-Core-Agent Harness for Production AI Agents","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expa-round2-article",[],{"intro":2407,"solving-context-coherence":3407,"anatomy-of-harness":3423,"gan-feedback-loop":3444,"production-patterns":3483,"economics-of-harness":3496,"conclusion":3354},{"title":3391,"description":3524},{"loc":3534},"3-core-agent-harness-production-architecture","articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expa-round2-article","CBd2pv8olaPjtCyjEO63oVaTjk18WRC-a7F-z0yiopo",{"id":3543,"title":3544,"ai_summary":3545,"author":8,"body":3546,"category":3918,"date_modified":601,"description":3550,"excerpt":8,"extension":603,"faq":601,"keywords":3919,"meta":3924,"meta_description":3925,"meta_title":3926,"navigation":616,"path":3927,"primary_keyword":3920,"published_at":601,"related_posts":3928,"sections":3929,"seo":3930,"sitemap":3931,"slug":3932,"status":623,"stem":3933,"__hash__":3934},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expb-premium-post-merge-article.md","Why Production AI Needs an Agent Harness, Not Just a Framework","An AI agent harness is a control plane that governs agent execution, enforces safety guardrails, and manages budgets. Unlike frameworks that orchestrate workflows, a harness uses a Planner, Generator, and Evaluator architecture to ensure reliable, cost-effective, and safe task completion in production environments.",{"type":10,"value":3547,"toc":3908},[3548,3551,3554,3557,3560,3564,3567,3570,3573,3576,3579,3582,3602,3611,3615,3618,3621,3624,3627,3636,3640,3643,3646,3649,3660,3663,3751,3754,3758,3761,3764,3767,3770,3773,3777,3780,3783,3786,3795,3801,3804,3812,3815,3818,3822,3825,3828,3831,3834,3837,3840,3866,3874,3876,3879,3882,3885,3888,3891,3905],[13,3549,3550],{},"An AI agent harness is a control plane that governs execution, enforces safety guardrails, and manages budgets. Unlike standard orchestration frameworks, a harness uses a Planner, Generator, and Evaluator architecture to guarantee reliable, cost-effective, and safe task completion in production environments.",[13,3552,3553],{},"You built a highly capable agent demo using a standard orchestration framework. It looks great on your laptop. But deploying it to production feels like letting a toddler drive a forklift. Without strict boundaries, agents drift. They hallucinate. They burn through API budgets in infinite reasoning loops.",[13,3555,3556],{},"Agent sprawl is the most common failure mode for teams moving from prototype to production. You wire together multiple models, give them tool access, and hope for the best. But generative models are inherently bad at grading their own homework.",[13,3558,3559],{},"Production reliability requires a different mental model. You have to treat agent orchestration as a control systems problem, not just a chaining problem.",[46,3561,3563],{"id":3562},"why-frameworks-arent-enough-for-production","Why Frameworks Aren't Enough for Production",[13,3565,3566],{},"Frameworks provide the building blocks for agent orchestration, but they lack built-in governance for safety, cost, and reliability. Relying solely on frameworks often leads to agent sprawl, where multiple loosely coordinated models duplicate costs and bypass compliance policies. Production systems require an independent control layer to prevent prototype drift and ensure consistent execution.",[13,3568,3569],{},"Frameworks like LangChain or CrewAI are excellent for wiring up inputs and outputs. They give you the primitives to connect a large language model to a vector database or a search API.",[13,3571,3572],{},"But they do not inherently restrict bad behavior. They assume the model will follow your system prompt.",[13,3574,3575],{},"Loose coordination feels fast during prototyping, but it creates massive technical debt. You end up with duplicated contexts across multiple agents. This drives up token costs unnecessarily and fragments your system's state.",[13,3577,3578],{},"Without a governing layer, long-running workflows fail silently. Worse, they get trapped in infinite reasoning loops. An agent might repeatedly call a failing API, burning through tokens until hitting a hard timeout.",[13,3580,3581],{},"Relying entirely on frameworks introduces three distinct failure modes:",[183,3583,3584,3590,3596],{},[186,3585,3586,3589],{},[636,3587,3588],{},"Missing financial limits:"," Agents lack awareness of token spend per session and will exhaust budgets to solve impossible tasks.",[186,3591,3592,3595],{},[636,3593,3594],{},"Unbounded execution:"," Workflows can stall in loops without an independent referee to terminate them.",[186,3597,3598,3601],{},[636,3599,3600],{},"Context duplication:"," Loosely coupled agents constantly re-read the same state, inflating operational costs.",[13,3603,3604,3605,3610],{},"According to ",[20,3606,3609],{"href":3607,"rel":3608},"https:\u002F\u002Fatlan.com\u002Fknow\u002Fbest-ai-agent-harness-tools-2026\u002F",[3063],"industry analysis by Atlan",", 80% of agentic AI implementation time is consumed by data engineering and governance, rather than framework configuration. No framework governs what agents actually read or execute. That is a structural problem you have to solve at the architectural level.",[46,3612,3614],{"id":3613},"what-is-an-ai-agent-harness","What Is an AI Agent Harness?",[13,3616,3617],{},"An AI agent harness acts as the control plane for your generative models, managing their lifecycle, tool access, and API budgets. Instead of just passing prompts, the harness enforces guardrails and execution limits. The industry standard relies on three core roles—a Planner, a Generator, and an Evaluator—to maintain strict oversight.",[13,3619,3620],{},"Think of the harness as the Kubernetes for your agents. It does not write the code or generate the text. It manages state, enforces policies, and handles the lifecycle of the models doing the work. It sits between your application logic and the LLM APIs.",[13,3622,3623],{},"A harness centralizes observability. When an agent loop fails, you need to trace exactly why it failed. A harness logs the inputs, outputs, tokens consumed, and tool calls at every step, giving you a clear audit trail.",[13,3625,3626],{},"It also operationalizes safety nets. If an agent attempts an unauthorized database drop, the harness intercepts and blocks the tool call before execution. The agent proposes the action; the harness decides whether to execute it.",[13,3628,3629,3630,3635],{},"The architecture driving this control plane is the Planner-Generator-Evaluator pattern. Research from Anthropic's engineering teams demonstrates that separating planning, generation, and evaluation enables better handling of subjective assessments while maintaining reproducibility ",[17,3631,3632],{},[20,3633,26],{"href":31,"ariaDescribedBy":3634,"dataFootnoteRef":8,"id":33},[24],". You stop asking one model to do everything.",[46,3637,3639],{"id":3638},"the-planner-defining-the-machine-checkable-contract","The Planner: Defining the Machine-Checkable Contract",[13,3641,3642],{},"The Planner agent breaks high-level goals into detailed, actionable specifications using highly capable reasoning models like Claude 3.5 Sonnet or GPT-4o. It translates ambiguous user requests into strict behavioral requirements and explicit verification assertions. This creates a machine-checkable contract that dictates exactly what the downstream models must produce to succeed.",[13,3644,3645],{},"The Planner requires the most context and domain knowledge. It consumes the user's prompt, retrieves necessary documentation, and figures out the architecture of the solution. It does not write the final implementation.",[13,3647,3648],{},"Output from the Planner must be heavily structured. You want JSON, not prose. The harness needs to parse this output programmatically to route instructions to the next agent in the loop.",[13,3650,3651,3652,3655,3656,3659],{},"A good plan includes explicit ",[270,3653,3654],{},"VERIFY"," assertions. These assertions act as a grading rubric for the rest of the system. If the task is to write a Python data scraper, the Planner should output assertions like \"Script must use the ",[270,3657,3658],{},"requests"," library\" and \"Script must include error handling for HTTP 404\".",[13,3661,3662],{},"Consider this example of a Planner's JSON output:",[264,3664,3666],{"className":1182,"code":3665,"language":1184,"meta":8,"style":8},"{\n  \"task\": \"Fetch user data and format as CSV\",\n  \"steps\": [\n    \"Initialize API client\", \n    \"Paginate through \u002Fusers\", \n    \"Write to output.csv\"\n  ],\n  \"assertions\": [\n    \"VERIFY: Uses pagination tokens\",\n    \"VERIFY: Creates output.csv in the current directory\",\n    \"VERIFY: Does not log PII to stdout\"\n  ]\n}\n",[270,3667,3668,3672,3684,3692,3700,3707,3712,3716,3723,3730,3737,3742,3747],{"__ignoreMap":8},[273,3669,3670],{"class":275,"line":276},[273,3671,1191],{"class":283},[273,3673,3674,3677,3679,3682],{"class":275,"line":287},[273,3675,3676],{"class":1024},"  \"task\"",[273,3678,296],{"class":283},[273,3680,3681],{"class":299},"\"Fetch user data and format as CSV\"",[273,3683,1226],{"class":283},[273,3685,3686,3689],{"class":275,"line":303},[273,3687,3688],{"class":1024},"  \"steps\"",[273,3690,3691],{"class":283},": [\n",[273,3693,3694,3697],{"class":275,"line":314},[273,3695,3696],{"class":299},"    \"Initialize API client\"",[273,3698,3699],{"class":283},", \n",[273,3701,3702,3705],{"class":275,"line":724},[273,3703,3704],{"class":299},"    \"Paginate through \u002Fusers\"",[273,3706,3699],{"class":283},[273,3708,3709],{"class":275,"line":739},[273,3710,3711],{"class":299},"    \"Write to output.csv\"\n",[273,3713,3714],{"class":275,"line":752},[273,3715,2021],{"class":283},[273,3717,3718,3721],{"class":275,"line":763},[273,3719,3720],{"class":1024},"  \"assertions\"",[273,3722,3691],{"class":283},[273,3724,3725,3728],{"class":275,"line":771},[273,3726,3727],{"class":299},"    \"VERIFY: Uses pagination tokens\"",[273,3729,1226],{"class":283},[273,3731,3732,3735],{"class":275,"line":781},[273,3733,3734],{"class":299},"    \"VERIFY: Creates output.csv in the current directory\"",[273,3736,1226],{"class":283},[273,3738,3739],{"class":275,"line":791},[273,3740,3741],{"class":299},"    \"VERIFY: Does not log PII to stdout\"\n",[273,3743,3744],{"class":275,"line":809},[273,3745,3746],{"class":283},"  ]\n",[273,3748,3749],{"class":275,"line":821},[273,3750,1245],{"class":283},[13,3752,3753],{},"This contract stops scope creep. The downstream models only see this constrained specification, keeping them focused entirely on execution. They do not need to know the broader business context, which saves massive amounts of context window space.",[46,3755,3757],{"id":3756},"the-generator-executing-fast-and-fluently","The Generator: Executing Fast and Fluently",[13,3759,3760],{},"The Generator agent focuses entirely on implementing the Planner's specifications by producing code, text, or tool actions. Because it does not need to plan or evaluate its own work, you can optimize this role for speed and cost by using specialized, highly fluent models to avoid getting stuck in reasoning loops.",[13,3762,3763],{},"The Generator is the blue-collar worker of the harness. It just writes the code or executes the API calls. It receives a highly constrained prompt containing only the current step's requirements.",[13,3765,3766],{},"This separation of concerns drastically reduces token overhead. You can route this step to faster, cheaper models. The cognitive heavy lifting was already done by the Planner, so the Generator only needs to be fluent in the target syntax.",[13,3768,3769],{},"Dynamic multi-model routing is a massive advantage here. You might use a heavy frontier model for planning, but rely on a lightweight, fast model like Claude 3.5 Haiku or GPT-4o-mini for the generation step. This cuts costs without sacrificing quality.",[13,3771,3772],{},"Stripping away the planning responsibility prevents the Generator from overthinking. When models try to plan, write, and critique simultaneously, they frequently stall. The Generator just executes the contract and returns the result to the harness.",[46,3774,3776],{"id":3775},"the-evaluator-independent-verification","The Evaluator: Independent Verification",[13,3778,3779],{},"The Evaluator agent independently assesses the Generator's output against the Planner's original specification without bias. Models reliably overrate their own outputs, making self-reflection highly error-prone in production.",[13,3781,3782],{},"By isolating the evaluation step with a separate model, the system generates structured, skeptical feedback to catch hallucinations before they reach the user.",[13,3784,3785],{},"The Evaluator acts like a senior engineer reviewing a pull request. It must be explicitly prompted to be skeptical. If you ask a model \"Is this good?\", it will almost always say yes.",[13,3787,3788,3789,3794],{},"Generative models exhibit severe self-evaluation bias. ",[20,3790,3793],{"href":3791,"rel":3792},"https:\u002F\u002Fwww.mindstudio.ai\u002Fblog\u002Fplanner-generator-evaluator-pattern-gan-inspired-ai-coding",[3063],"Industry patterns documented by MindStudio"," emphasize that the Evaluator must provide structured feedback the Generator can act on.",[13,3796,3797,3798,3800],{},"Instead of just a binary pass\u002Ffail signal, it needs to point out exactly which ",[270,3799,3654],{}," assertion failed and why.",[13,3802,3803],{},"You must strictly isolate the Evaluator. Separating this role prevents the Generator from grading its own homework.",[13,3805,3806,3811],{},[20,3807,3810],{"href":3808,"rel":3809},"https:\u002F\u002Fredis.io\u002Fblog\u002Fai-agent-architecture-patterns\u002F",[3063],"Analysis from Redis"," highlights that agent performance degrades significantly over multiple consecutive runs without proper architecture and evaluation discipline.",[13,3813,3814],{},"The Evaluator does not fix the code. It returns a critique document.",[13,3816,3817],{},"The harness reads this document. If the critique contains failures, the harness routes the feedback back to the Generator for another attempt.",[46,3819,3821],{"id":3820},"implementing-your-ai-agent-harness-loop","Implementing Your AI Agent Harness Loop",[13,3823,3824],{},"Building an AI agent harness loop requires strict termination conditions, such as evaluator pass signals, maximum iteration caps, or confidence thresholds. The harness executes the Generator and Evaluator in a cycle until the acceptance criteria are met or the budget is exhausted, guaranteeing incremental progress while maintaining strict observability.",[13,3826,3827],{},"You have to set hard limits on iterations. A common standard is capping the loop at three attempts. If the Generator cannot pass the Evaluator after three tries, the harness must terminate the loop to prevent runaway costs.",[13,3829,3830],{},"Log every generation and evaluation pair. When you review the logs, you will quickly identify which prompts or steps consistently fail.",[13,3832,3833],{},"This data is critical for refining your Planner's initial specifications over time.",[13,3835,3836],{},"For high-stakes operations, implement human-in-the-loop escalation. If the Evaluator repeatedly rejects the Generator's output, or if the confidence threshold drops, the harness pauses execution. It then routes the state to a human operator for manual review.",[13,3838,3839],{},"A production loop generally follows these stages:",[183,3841,3842,3848,3854,3860],{},[186,3843,3844,3847],{},[636,3845,3846],{},"Plan:"," The harness calls the Planner to generate the JSON specification and assertions.",[186,3849,3850,3853],{},[636,3851,3852],{},"Generate:"," The harness passes the spec to the Generator to execute step one.",[186,3855,3856,3859],{},[636,3857,3858],{},"Evaluate:"," The harness passes the output and the assertions to the Evaluator.",[186,3861,3862,3865],{},[636,3863,3864],{},"Route:"," The harness reads the Evaluator's JSON response. On failure, it loops back to Generate. On success, it proceeds to the next step.",[13,3867,3868,3873],{},[20,3869,3872],{"href":3870,"rel":3871},"https:\u002F\u002Fdev.to\u002Ffmquaglia\u002Fbeyond-the-notebook-4-architectural-patterns-for-production-ready-ai-agents-3a16",[3063],"Architectural patterns for production-ready AI agents"," require treating the entire loop as a state machine. The harness holds the state. The agents are just stateless functions called by the harness.",[46,3875,3354],{"id":3353},[13,3877,3878],{},"Moving AI from a notebook demo to a production system requires abandoning the idea of a single omnipotent agent. Instead, builders must adopt a strictly governed Planner, Generator, and Evaluator architecture to ensure reliability, control costs, and maintain safety at scale.",[13,3880,3881],{},"Frameworks will help you build the components, but the harness is what keeps them running safely. You have to own the control loop. Relying on an LLM to police its own execution is an unforced error.",[13,3883,3884],{},"Open your current agent workflow code and identify where planning, generation, and validation happen in the same prompt. Extract the validation step into a separate, strongly typed model call to immediately improve reliability.",[13,3886,3887],{},"Ready to stop agent drift and build resilient systems? Start implementing an independent evaluation layer today, and subscribe to our newsletter for more deep dives into production AI architecture.",[3889,3890],"hr",{},[391,3892,3894,3897],{"className":3893,"dataFootnotes":8},[394],[46,3895,399],{"className":3896,"id":24},[398],[401,3898,3899],{},[186,3900,3901,3902],{"id":460},"InfoQ. \"Anthropic Designs Three-Agent Harness Supports Long-Running Full-Stack AI Development.\" ",[20,3903,413],{"href":464,"ariaLabel":410,"className":3904,"dataFootnoteBackref":8},[412],[588,3906,3907],{},"html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":8,"searchDepth":287,"depth":287,"links":3909},[3910,3911,3912,3913,3914,3915,3916,3917],{"id":3562,"depth":287,"text":3563},{"id":3613,"depth":287,"text":3614},{"id":3638,"depth":287,"text":3639},{"id":3756,"depth":287,"text":3757},{"id":3775,"depth":287,"text":3776},{"id":3820,"depth":287,"text":3821},{"id":3353,"depth":287,"text":3354},{"id":24,"depth":287,"text":399},"ai-engineering",[3920,3208,3921,3922,3923,609],"AI agent harness","agentic architecture","AI control plane","LLM evaluation",{},"Moving AI to production requires strict governance. Discover how an AI agent harness uses a Planner, Generator, and Evaluator to ensure safe execution.","AI Agent Harness: The Planner-Generator-Evaluator Architecture","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expb-premium-post-merge-article",[],{"intro":2407,"why-frameworks-arent-enough-for-production":3563,"what-is-an-ai-agent-harness":3614,"the-planner-defining-the-machine-checkable-contract":3639,"the-generator-executing-fast-and-fluently":3757,"the-evaluator-independent-verification":3776,"implementing-your-ai-agent-harness-loop":3821,"conclusion":3354},{"title":3544,"description":3550},{"loc":3927},"ai-agent-harness-production-architecture","articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expb-premium-post-merge-article","n4YrqKIbgDFFExlRCdf0McqYftoHNMD9Ps39zY0tMA4",{"id":3936,"title":3937,"ai_summary":3938,"author":8,"body":3939,"category":3369,"date_modified":601,"description":3938,"excerpt":8,"extension":603,"faq":601,"keywords":4128,"meta":4134,"meta_description":4135,"meta_title":4136,"navigation":616,"path":4137,"primary_keyword":4129,"published_at":601,"related_posts":4138,"sections":4139,"seo":4140,"sitemap":4141,"slug":4142,"status":623,"stem":4143,"__hash__":4144},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expb-round2-article.md","Building Reliable AI with the Planner-Generator-Evaluator Pattern","The planner-generator-evaluator pattern is a multi-agent architecture that separates reasoning, execution, and quality assurance into distinct roles. By forcing an adversarial feedback loop between generation and skeptical evaluation, this harness eliminates the self-confirmation bias and hallucination loops that plague single-agent systems.",{"type":10,"value":3940,"toc":4119},[3941,3943,3946,3949,3952,3956,3959,3962,3965,3968,3972,3975,3981,3986,3996,4000,4003,4010,4013,4022,4026,4029,4032,4052,4055,4058,4067,4071,4074,4077,4086,4094,4097,4099,4102,4105],[13,3942,3938],{},[13,3944,3945],{},"Most AI agents look great in a controlled social media demo. They write a quick Python script or draft an email, and the output seems flawless on the first try. But these same monolithic agents fail spectacularly when asked to write production code, validate infrastructure setups, or manage a multi-step workflow without human intervention.",[13,3947,3948],{},"The core problem is scaling AI from single-shot prompts to reliable, long-running production tasks. When a single foundation model tries to plan a complex task, execute the steps, and check its own work, it inevitably cuts corners. It forgets initial constraints, hallucinates variables, and rubber-stamps its own errors.",[13,3950,3951],{},"The planner-generator-evaluator (PGE) harness offers an architectural blueprint to fix this, avoiding the trap of generic, bloated multi-agent frameworks. It enforces strict separation of concerns across the lifecycle of a task. Engineering teams must stop relying on single models to grade their own homework. To ship reliable AI features, you need structural constraints that force models to prove their work.",[46,3953,3955],{"id":3954},"the-fallacy-of-single-agent-systems","The Fallacy of Single-Agent Systems",[13,3957,3958],{},"Single-agent systems fail at long-running tasks because generation and evaluation share the same context window, creating an inescapable self-confirmation bias. When a model drafts code and immediately reviews it, it inherently trusts its own logic, blinding it to edge-case failures and gradual coherence loss.",[13,3960,3961],{},"When you ask a large language model (LLM) to generate a complex script and then follow up with \"is this correct?\", the model reads its own output in the context window. It recognizes the exact probabilistic patterns it just generated, and because those patterns match its internal weights, it assumes the output is correct. This self-confirmation bias makes single agents terrible at catching their own logical flaws, especially in strict domains like continuous integration pipelines.",[13,3963,3964],{},"As the context window fills with previous steps, the model's reasoning capabilities degrade further. It begins to lose track of the original constraints. If you ask a single agent to generate a React component, add unit tests, and then fix accessibility issues, by the third step it has likely forgotten the styling rules you established in step one. The attention mechanism spreads too thin across the growing token history, leading to hallucinated variables and skipped requirements.",[13,3966,3967],{},"The common industry reaction to these failures is to reach for off-the-shelf multi-agent frameworks. These generic tools often add massive coordination overhead without structurally solving the feedback loop problem. They give you a dozen agents chatting with each other, but no mathematical or architectural guarantee of quality. Reliable AI requires deliberate architectural constraints, not just larger foundation models or noisier agent swarms.",[46,3969,3971],{"id":3970},"deconstructing-the-planner-generator-evaluator-pattern","Deconstructing the Planner-Generator-Evaluator Pattern",[13,3973,3974],{},"The planner-generator-evaluator pattern divides complex AI tasks into three specialized roles to enforce strict quality control and prevent context contamination. The Planner expands a high-level prompt into a rigorous specification, the Generator produces the required output against that spec, and the Evaluator independently critiques the result until all criteria are met.",[13,3976,3977,3978,3980],{},"This architecture explicitly isolates reasoning from execution. The ",[636,3979,1887],{}," role is responsible for scoping edge cases, defining strict acceptance criteria, and setting an ambitious bar for the output. Instead of writing the final code or content, it writes the blueprint by translating a vague user request into a deterministic checklist. By forcing the system to declare its intentions before writing code, the Planner establishes a ground-truth document that anchors the rest of the workflow.",[13,3982,3429,3983,3985],{},[636,3984,1495],{}," role operates in a narrow, highly focused context. It receives the Planner's specification and executes it without the distraction of higher-level reasoning. By stripping away the conversational history and the original user prompt, the Generator can dedicate its entire attention mechanism to satisfying the exact constraints of the spec. It functions purely as a translation layer between the specification and the final syntax.",[13,3987,3429,3988,3990,3991,3995],{},[636,3989,1501],{}," role acts as an independent quality assurance mechanism. It provides structured, granular feedback rather than a simple binary pass\u002Ffail. This setup adapts the adversarial loop of Generative Adversarial Networks (GANs) for concrete task completion. Instead of training a model's weights, this ",[20,3992,3994],{"href":3791,"rel":3993},[3063],"GAN-inspired architecture"," iterates on a specific output through rigorous, automated critique. The Evaluator compares the Generator's draft directly against the Planner's exact criteria, sending failures back for correction.",[46,3997,3999],{"id":3998},"why-the-evaluator-role-dictates-success","Why the Evaluator Role Dictates Success",[13,4001,4002],{},"A production-grade Evaluator operates as a skeptical quality assurance engineer rather than a passive reviewer. By utilizing a clean context window to judge the Generator's output strictly against the Planner's original specification, this adversarial stance prevents rubber-stamping and forces iterative improvement cycles until the output legitimately passes all constraints.",[13,4004,4005,4006,4009],{},"Context isolation is the most critical factor in this phase. The Evaluator must never see the Generator's internal reasoning or the conversational history that led to the draft. It should only receive the final output and the initial specification. If the Evaluator sees the Generator explaining ",[3347,4007,4008],{},"why"," it made a mistake, the Evaluator is more likely to forgive the error. Maintaining a strict firewall between these contexts forces the Evaluator to judge the work purely on its objective merits.",[13,4011,4012],{},"Prompt engineering tactics for the Evaluator must actively enforce skepticism. You have to explicitly instruct the model to assume the output is flawed. The prompt should demand that the Evaluator search for specific failure modes, edge cases, and missing requirements. It must output a structured assessment—often a JSON payload detailing exact line numbers and missing criteria—rather than a conversational summary.",[13,4014,4015,4016,4021],{},"Top engineering teams combine this static LLM review with dynamic runtime validation. For example, ",[20,4017,4020],{"href":4018,"rel":4019},"https:\u002F\u002Fwww.shiplight.ai\u002Fblog\u002Fplanner-generator-evaluator-multi-agent-qa",[3063],"OpenAI wired the Chrome DevTools Protocol into their agent runtime"," to validate UI fixes in a real browser. This functional observation provides a ground-truth signal that static code inspection cannot match. When the Evaluator finds a flaw, its detailed, structured feedback becomes the exact prompt for the Generator's next iteration.",[46,4023,4025],{"id":4024},"implementing-the-planner-generator-evaluator-pattern-in-production","Implementing the Planner-Generator-Evaluator Pattern in Production",[13,4027,4028],{},"Implementing the planner-generator-evaluator pattern in production requires assigning distinct foundation models to each role based on their strengths. This balances reasoning capabilities with token costs. Teams route complex planning and evaluation to frontier models, assigning repetitive generation to faster, cost-effective alternatives.",[13,4030,4031],{},"You do not need to use the most expensive model for every step. A standard routing strategy looks like this:",[183,4033,4034,4040,4046],{},[186,4035,4036,4039],{},[636,4037,4038],{},"Planner:"," Claude 3.5 Sonnet or GPT-4o (high reasoning, deep context)",[186,4041,4042,4045],{},[636,4043,4044],{},"Generator:"," Claude 3.5 Haiku or GPT-4o-mini (fast execution, low cost)",[186,4047,4048,4051],{},[636,4049,4050],{},"Evaluator:"," Claude 3.5 Sonnet or GPT-4o (strict logic, critical analysis)",[13,4053,4054],{},"By deliberately splitting these tasks, you optimize for both budget and performance without degrading the final output.",[13,4056,4057],{},"You must define strict iteration limits to prevent infinite loops and control API spend. Cap the evaluation loop at three to five cycles. If the Generator cannot satisfy the Evaluator after five attempts, the system should halt and escalate the task to a human developer. A loop that fails to converge usually indicates a flawed initial plan or an impossible specification, not a generation error.",[13,4059,4060,4061,4066],{},"Orchestration requires managing state, scheduling, and communication between these isolated contexts. You can wire these agents together using code-first SDKs like ",[20,4062,4065],{"href":4063,"rel":4064},"https:\u002F\u002Fchatbotkit.com\u002Fexamples\u002Fplanner-executor-evaluator-harness",[3063],"ChatBotKit"," or visual workflow builders. Regardless of the tooling, keep the system prompts entirely isolated per role to maintain the integrity of the adversarial harness. State management must rigorously track which iteration cycle the system is currently executing to enforce the termination thresholds.",[46,4068,4070],{"id":4069},"the-cost-and-complexity-trade-offs","The Cost and Complexity Trade-Offs",[13,4072,4073],{},"Adopting a three-agent harness introduces significant architectural complexity, increasing token overhead, latency, and debugging difficulty. Engineering teams must earn this complexity by proving a single-agent baseline cannot meet the required quality bar, investing heavily in observability tooling to trace which agent introduced an error.",[13,4075,4076],{},"Multi-agent loops cost more and run slower. Passing context back and forth between three distinct models over multiple network requests adds unavoidable latency. If your user expects a real-time response, an iterative evaluation loop will feel sluggish and unresponsive. You are fundamentally trading speed for mathematical correctness and reliability.",[13,4078,4079,4080,4085],{},"Debugging also becomes a distributed systems problem. A failure could stem from a vague plan, a hallucinated generation, or an overly pedantic evaluation. As noted in the ",[20,4081,4084],{"href":4082,"rel":4083},"https:\u002F\u002Fwww.codebridge.tech\u002Farticles\u002Fsingle-agent-vs-multi-agent-architecture-what-changes-in-reliability-cost-and-debuggability",[3063],"Codebridge decision framework",", teams that ship reliable AI systems are those who earned every agent they added. You must be able to isolate whether the Planner missed a constraint or the Evaluator triggered a false positive.",[13,4087,4088,4089],{},"Do not use this pattern for simple text summarization or basic data extraction. Reserve the PGE harness for high-stakes domains like code generation, continuous integration configuration, or infrastructure as code.",[17,4090,4091],{},[20,4092,26],{"href":22,"ariaDescribedBy":4093,"dataFootnoteRef":8,"id":25},[24],[13,4095,4096],{},"To maintain system observability, you must mandate robust logging for every iteration cycle. Record the exact prompt, the raw output, the token count, and the latency for every single role in the loop. Without this granular telemetry, diagnosing a failed multi-agent workflow devolves into blind guesswork.",[46,4098,3354],{"id":3353},[13,4100,4101],{},"The three-agent harness proves that reliable AI outputs require architectural discipline and strict separation of concerns. Moving beyond monolithic prompts into specialized, adversarial feedback loops catches errors before they reach production. It forces foundation models to operate within the same quality assurance constraints we expect from human engineering teams.",[13,4103,4104],{},"Audit your most error-prone single-agent workflow today. Split the existing prompt into a distinct planning phase and a generation phase, and manually act as the Evaluator yourself to measure the quality delta. Do this before you write any orchestration code.",[391,4106,4108,4111],{"className":4107,"dataFootnotes":8},[394],[46,4109,399],{"className":4110,"id":24},[398],[401,4112,4113],{},[186,4114,4115,4116],{"id":405},"Anthropic engineering notes that harness design for long-running apps demands continuous iteration and monitoring. It is a practice, not a pattern you implement once and forget. ",[20,4117,413],{"href":409,"ariaLabel":410,"className":4118,"dataFootnoteBackref":8},[412],{"title":8,"searchDepth":287,"depth":287,"links":4120},[4121,4122,4123,4124,4125,4126,4127],{"id":3954,"depth":287,"text":3955},{"id":3970,"depth":287,"text":3971},{"id":3998,"depth":287,"text":3999},{"id":4024,"depth":287,"text":4025},{"id":4069,"depth":287,"text":4070},{"id":3353,"depth":287,"text":3354},{"id":24,"depth":287,"text":399},[4129,4130,4131,4132,4133,3373],"planner-generator-evaluator pattern","multi-agent architecture","AI agent production","LLM orchestration","AI evaluation loop",{},"Learn how the planner-generator-evaluator pattern fixes single-agent hallucinations by separating reasoning, execution, and QA into distinct roles.","The Planner-Generator-Evaluator Pattern for AI Agents","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expb-round2-article",[],{"intro":2407,"the-fallacy-of-single-agent-systems":3955,"deconstructing-the-planner-generator-evaluator-pattern":3971,"why-the-evaluator-role-dictates-success":3999,"implementing-the-planner-generator-evaluator-pattern-in-production":4025,"the-cost-and-complexity-trade-offs":4070},{"title":3937,"description":3938},{"loc":4137},"planner-generator-evaluator-pattern","articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-ca-203-expb-round2-article","LnDdAcnBrlLUDdhOLM6KJwu-rn73rrBJBoIvD13uA-c",{"id":4146,"title":4147,"ai_summary":4148,"author":8,"body":4149,"category":1779,"date_modified":601,"description":5242,"excerpt":8,"extension":603,"faq":601,"keywords":5243,"meta":5249,"meta_description":5250,"meta_title":5251,"navigation":616,"path":5252,"primary_keyword":3371,"published_at":601,"related_posts":5253,"sections":5254,"seo":5255,"sitemap":5256,"slug":5257,"status":623,"stem":5258,"__hash__":5259},"articles\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-e1-budget-8k-article.md","The 3-Core-Agent Harness: Why Production Agent Systems Need Planner + Generator + Evaluator, Not Frameworks","Production agent systems use a 3-core-agent harness: Planner sets high-level specs, Generator builds complete outputs autonomously, Evaluator scores with adversarial rubrics. This ditches framework bloat, leverages Claude Opus 4.6's 1M-token coherence for reliable long-horizon tasks without error cascades or oversight.",{"type":10,"value":4150,"toc":5231},[4151,4162,4171,4175,4183,4192,4200,4203,4228,4236,4244,4247,4256,4259,4264,4267,4282,4288,4291,4295,4304,4314,4328,4334,4337,4341,4350,4359,4365,4374,4378,4387,4397,4407,4422,4425,4431,4434,4629,4632,4641,4644,4704,4707,4710,4714,4722,4731,4802,4812,4816,4825,4827,4844,4847,4910,4921,4925,4931,4947,4950,5012,5015,5024,5228],[13,4152,4153,4154,4156,4157],{},"Production agent systems thrive with a ",[636,4155,3371],{},"—Planner for high-level specs, Generator for implementation, Evaluator for rigorous checks. Ditch bloated frameworks. Modern LLMs like Claude Opus 4.6 use 1M-token contexts and coherence for reliable, scalable outputs on complex tasks.",[17,4158,4159],{},[20,4160,26],{"href":22,"ariaDescribedBy":4161,"dataFootnoteRef":8,"id":25},[24],[13,4163,4164,4165,4170],{},"Anthropic's leaked code and experiments reveal that 90% of components in frameworks like BMAD, GSD, and SpecKit add overhead without boosting long-horizon success.",[17,4166,4167],{},[20,4168,26],{"href":22,"ariaDescribedBy":4169,"dataFootnoteRef":8,"id":43},[24]," Builders waste cycles on error-prone sharding designed for weak models. This harness cuts dev time, reduces bugs, and ships trustworthy AI—test it by stripping your stack today.",[46,4172,4174],{"id":4173},"why-agent-frameworks-fail-production-systems","Why Agent Frameworks Fail Production Systems",[13,4176,4177,4178],{},"Agent frameworks like BMAD, GSD, SpecKit, and Superpowers fail production systems because they over-engineer solutions designed for outdated LLM limits—such as short context windows, mid-task hallucinations, and poor coherence. With Claude Opus 4.6's 1M-token window, micro-task sharding, frequent resets, and sub-agent handoffs create unnecessary overhead, amplify error cascades through rigid chains, and deliver worse long-horizon performance than stripped-down 3-core harnesses.",[17,4179,4180],{},[20,4181,26],{"href":22,"ariaDescribedBy":4182,"dataFootnoteRef":8,"id":58},[24],[13,4184,4185,4186,4191],{},"Anthropic's internal tests stripped these frameworks layer by layer. They found 90% of components delivered zero value on complex tasks; removing them actually improved success rates.",[17,4187,4188],{},[20,4189,26],{"href":22,"ariaDescribedBy":4190,"dataFootnoteRef":8,"id":67},[24]," Detailed specs lock agents into early flaws, while self-evaluation misses subtle bugs, especially in UI flows.",[13,4193,4194,4195],{},"Frameworks made sense for short-context models that hallucinated mid-task. Now, with coherence spanning entire projects, they bloat pipelines unnecessarily.",[17,4196,4197],{},[20,4198,34],{"href":31,"ariaDescribedBy":4199,"dataFootnoteRef":8,"id":33},[24],[13,4201,4202],{},"Common pitfalls include overconfident pass\u002Ffail checks and rigid sharding that propagates one bad step across the chain. I've seen teams burn weeks debugging these cascades—simpler harnesses fix that.",[183,4204,4205,4211,4217],{},[186,4206,4207,4210],{},[636,4208,4209],{},"Error propagation",": Micro-tasks amplify a single planner mistake into full failures.",[186,4212,4213,4216],{},[636,4214,4215],{},"Overhead costs",": Extra agents and resets double token spend without gains.",[186,4218,4219,4222,4223],{},[636,4220,4221],{},"UI blind spots",": Self-eval rates flawed designs as perfect 80% of the time.",[17,4224,4225],{},[20,4226,26],{"href":22,"ariaDescribedBy":4227,"dataFootnoteRef":8,"id":92},[24],[13,4229,4230,4231],{},"Anthropic's leaked experiments provide a concrete case study: building a multi-step e-commerce checkout flow. Frameworks like BMAD and GSD shard it into 15+ micro-tasks—auth planning, cart UI design, payment integration, error handling—each with its own context reset. A flaw in the early cart shard (e.g., overlooked mobile responsiveness) cascades downstream, tanking 80% of full runs due to accumulated state loss.",[17,4232,4233],{},[20,4234,26],{"href":22,"ariaDescribedBy":4235,"dataFootnoteRef":8,"id":219},[24],[13,4237,4238,4239],{},"Stripping to a 3-core harness changed this. The Planner outputs high-level stories (\"secure, responsive checkout supporting Stripe and Apple Pay\"). The Generator builds the entire flow end-to-end in one 400k-token pass. Evaluator simulates 10 user journeys, catching issues like cart persistence fails. Success rates climbed as layers vanished, proving frameworks patch weak models but hobble strong ones.",[17,4240,4241],{},[20,4242,26],{"href":22,"ariaDescribedBy":4243,"dataFootnoteRef":8,"id":383},[24],[13,4245,4246],{},"I've replicated this locally on a task tracker app. GSD's sharding took 12 iterations with 40% failure from prop errors; harness nailed it in 3 loops at 90% success. Token spend dropped 35%. Builders, audit your stack: remove one layer today and measure.",[13,4248,4249,4250,4255],{},"Prior work on Anthropic's 3 agents showed frameworks bloat strong models while patching weak ones.",[17,4251,4252],{},[20,4253,79],{"href":76,"ariaDescribedBy":4254,"dataFootnoteRef":8,"id":78},[24]," The fix? Core roles only.",[46,4257,4258],{"id":1860},"What is the 3-Core-Agent Harness?",[13,4260,3429,4261,4263],{},[636,4262,3371],{}," structures production systems around Planner (high-level product outlines), Generator (autonomous implementation), and Evaluator (adversarial rubric-based critique)—replacing framework bloat to let advanced LLMs like Claude Opus discover optimal paths and deliver complete workflows reliably.",[13,4265,4266],{},"The loop runs iteratively: Plan → Generate → Evaluate. No sub-agents, no resets—just high-level scopes handed to a generator that builds end-to-end, then critiqued harshly. This mirrors real teams: PM outlines, engineer codes, QA tears it apart.",[13,4268,4269,4270,4275,4276,4281],{},"It outperforms frameworks on Anthropic's benchmarks because high-level plans avoid early locks, while separate eval catches generator oversights.",[17,4271,4272],{},[20,4273,26],{"href":22,"ariaDescribedBy":4274,"dataFootnoteRef":8,"id":1098},[24]," Aligns perfectly with the agent blueprint: Role + Goal + Tools + Rules + Output.",[17,4277,4278],{},[20,4279,108],{"href":105,"ariaDescribedBy":4280,"dataFootnoteRef":8,"id":107},[24]," Here's a simple flow:",[264,4283,4286],{"className":4284,"code":4285,"language":962},[960],"High-Level Spec ──→ Planner ──→ Outline\n                          │\n                          ↓\n                     Generator ──→ Full Output (Code\u002FUI)\n                          │\n                          ↓\n                     Evaluator ──→ Rubric Score + Fixes\n                          │\n                     Iterate until Pass\n",[270,4287,4285],{"__ignoreMap":8},[13,4289,4290],{},"Bloated frameworks look like spiderwebs by comparison—20+ nodes for what three do better.",[46,4292,4294],{"id":4293},"the-planner-defining-high-level-objectives","The Planner: Defining High-Level Objectives",[13,4296,4297,4298],{},"The Planner creates high-level product deliverables such as feature breakdowns, user stories, and phased rollouts instead of granular micro-tasks. This lets LLMs independently discover optimal implementation paths while preventing a single flaw in detailed specs from cascading into full workflow failures, as shown in Anthropic's framework stripping experiments.",[17,4299,4300],{},[20,4301,26],{"href":22,"ariaDescribedBy":4302,"dataFootnoteRef":8,"id":4303},[24],"user-content-fnref-1-9",[13,4305,4306,4307,4313],{},"Feed it a boundary-tested app idea: \"Build a task manager with user auth, drag-drop boards, and notifications.\" It spits phased docs, folders, and stories—no tech sharding.",[17,4308,4309],{},[20,4310,26],{"href":22,"ariaDescribedBy":4311,"dataFootnoteRef":8,"id":4312},[24],"user-content-fnref-1-10"," Anthropic's prompt example: High-level phases over BMAD's granular splits.",[13,4315,4316,4317,4322,4323],{},"Avoid native Claude plan mode; it dives too deep, risking early flaws. Tools shine here: BMAD for PRD gen, Superpowers for questioning assumptions.",[17,4318,4319],{},[20,4320,146],{"href":143,"ariaDescribedBy":4321,"dataFootnoteRef":8,"id":145},[24]," Extends Archon V3's YAML nodes for structured outputs.",[17,4324,4325],{},[20,4326,343],{"href":340,"ariaDescribedBy":4327,"dataFootnoteRef":8,"id":342},[24],[264,4329,4332],{"className":4330,"code":4331,"language":962},[960],"Prompt: \"As Product Manager, outline [idea] into deliverables: user stories, folder structure, phased rollout. Stay high-level—no code.\"\nOutput: YAML with epics, acceptance criteria, risks.\n",[270,4333,4331],{"__ignoreMap":8},[13,4335,4336],{},"This keeps generators flexible. One bad micro-spec? Whole chain tanks. High-level? They route around it.",[46,4338,4340],{"id":4339},"the-generator-implementation-without-micro-management","The Generator: Implementation Without Micro-Management",[13,4342,4343,4344],{},"The Generator receives the Planner's high-level spec and produces complete deliverables—codebases, UIs, content—managing implementation details independently using large context windows like Claude Opus 4.6's 1M tokens. It skips step-by-step hand-holding from frameworks designed for weaker models, enabling end-to-end workflows that reduce error propagation and handle dynamic paths effectively.",[17,4345,4346],{},[20,4347,26],{"href":22,"ariaDescribedBy":4348,"dataFootnoteRef":8,"id":4349},[24],"user-content-fnref-1-11",[13,4351,4352,4353,4358],{},"It leverages LLM coherence for end-to-end workflows: From outline to full app, Git commits included. Pairs with agent blueprint—no self-eval, just hand off.",[17,4354,4355],{},[20,4356,108],{"href":105,"ariaDescribedBy":4357,"dataFootnoteRef":8,"id":117},[24]," Example: Product spec → React app with Tailwind, tests, deploy script.",[264,4360,4363],{"className":4361,"code":4362,"language":962},[960],"Spec: \"Task manager with auth and boards.\"\nGenerator: Full repo—components, hooks, Prisma schema, Vercel config.\n",[270,4364,4362],{"__ignoreMap":8},[13,4366,4367,4368,4373],{},"Trade-off: Dynamic paths cost more tokens than chains but handle unknowns frameworks can't. Git integration prevents drift: Push to worktrees, PR-ready.",[17,4369,4370],{},[20,4371,343],{"href":340,"ariaDescribedBy":4372,"dataFootnoteRef":8,"id":369},[24]," Skips framework micromanagement since Claude holds 1M tokens flawlessly.",[46,4375,4377],{"id":4376},"the-evaluator-adversarial-checks-with-graded-rubrics","The Evaluator: Adversarial Checks with Graded Rubrics",[13,4379,4380,4381],{},"The Evaluator acts as an adversary by simulating user behaviors, hunting for bugs, and scoring outputs on weighted rubrics across UI axes like design consistency, originality, craftsmanship, and functionality—outperforming generator self-checks or binary pass\u002Ffail systems through detailed, iterative feedback for production-grade results.",[17,4382,4383],{},[20,4384,26],{"href":22,"ariaDescribedBy":4385,"dataFootnoteRef":8,"id":4386},[24],"user-content-fnref-1-12",[13,4388,4389,4390,4396],{},"Score on 4 UI axes, 1-10 each: Design (consistency), Originality (fresh UX), Craft (polish), Functionality (edge cases). Total weighted, require 8+ average.",[17,4391,4392],{},[20,4393,26],{"href":22,"ariaDescribedBy":4394,"dataFootnoteRef":8,"id":4395},[24],"user-content-fnref-1-13"," Playwright for live tests: \"Click login, drag task—does it crash?\"",[13,4398,4399,4400,4406],{},"Generators overlook 70% of their flaws; separate eval catches them.",[17,4401,4402],{},[20,4403,26],{"href":22,"ariaDescribedBy":4404,"dataFootnoteRef":8,"id":4405},[24],"user-content-fnref-1-14"," Frameworks like GSD approximate but skip grading.",[13,4408,4409,4410,4415,4416],{},"Archon V3 stats show this setup enables 3.5 PRs per engineer per day on million-line projects.",[17,4411,4412],{},[20,4413,343],{"href":340,"ariaDescribedBy":4414,"dataFootnoteRef":8,"id":376},[24]," Post-tool hooks add self-correction, per Archon V3.",[17,4417,4418],{},[20,4419,343],{"href":340,"ariaDescribedBy":4420,"dataFootnoteRef":8,"id":4421},[24],"user-content-fnref-6-4",[13,4423,4424],{},"To implement, feed the Evaluator a structured prompt:",[264,4426,4429],{"className":4427,"code":4428,"language":962},[960],"Prompt: \"Act as adversarial QA expert. Review the generated [output]. Simulate 5 diverse user sessions (mobile\u002Fdesktop, edge cases). Score 1-10 on each axis using this rubric. If weighted average \u003C8, list prioritized fixes with code diffs. Axes: Design (25%: Figma alignment, responsive), Originality (20%: unique UX vs templates), Craft (25%: polish, no errors), Functionality (30%: tests pass, no regressions).\"\n",[270,4430,4428],{"__ignoreMap":8},[13,4432,4433],{},"Integrate Playwright for automated verification:",[264,4435,4439],{"className":4436,"code":4437,"language":4438,"meta":8,"style":8},"language-javascript shiki shiki-themes github-light github-dark","\u002F\u002F evaluator_test.js (run via node)\nconst { test, expect } = require('@playwright\u002Ftest');\n\ntest('Full task manager flow', async ({ page }) => {\n  await page.goto('http:\u002F\u002Flocalhost:3000');\n  await page.fill('#email', 'user@test.com');\n  await page.click('#login');\n  await page.dragAndDrop('.task-item', '.target-board');\n  await expect(page.locator('.success-toast')).toBeVisible();\n  \u002F\u002F Flag failures: screenshots, console logs\n});\n","javascript",[270,4440,4441,4447,4482,4486,4516,4534,4555,4571,4592,4619,4624],{"__ignoreMap":8},[273,4442,4443],{"class":275,"line":276},[273,4444,4446],{"class":4445},"sJ8bj","\u002F\u002F evaluator_test.js (run via node)\n",[273,4448,4449,4453,4456,4459,4461,4464,4467,4470,4473,4476,4479],{"class":275,"line":287},[273,4450,4452],{"class":4451},"szBVR","const",[273,4454,4455],{"class":283}," { ",[273,4457,4458],{"class":1024},"test",[273,4460,801],{"class":283},[273,4462,4463],{"class":1024},"expect",[273,4465,4466],{"class":283}," } ",[273,4468,4469],{"class":4451},"=",[273,4471,4472],{"class":1020}," require",[273,4474,4475],{"class":283},"(",[273,4477,4478],{"class":299},"'@playwright\u002Ftest'",[273,4480,4481],{"class":283},");\n",[273,4483,4484],{"class":275,"line":303},[273,4485,1971],{"emptyLinePlaceholder":616},[273,4487,4488,4490,4492,4495,4497,4500,4503,4507,4510,4513],{"class":275,"line":314},[273,4489,4458],{"class":1020},[273,4491,4475],{"class":283},[273,4493,4494],{"class":299},"'Full task manager flow'",[273,4496,801],{"class":283},[273,4498,4499],{"class":4451},"async",[273,4501,4502],{"class":283}," ({ ",[273,4504,4506],{"class":4505},"s4XuR","page",[273,4508,4509],{"class":283}," }) ",[273,4511,4512],{"class":4451},"=>",[273,4514,4515],{"class":283}," {\n",[273,4517,4518,4521,4524,4527,4529,4532],{"class":275,"line":724},[273,4519,4520],{"class":4451},"  await",[273,4522,4523],{"class":283}," page.",[273,4525,4526],{"class":1020},"goto",[273,4528,4475],{"class":283},[273,4530,4531],{"class":299},"'http:\u002F\u002Flocalhost:3000'",[273,4533,4481],{"class":283},[273,4535,4536,4538,4540,4543,4545,4548,4550,4553],{"class":275,"line":739},[273,4537,4520],{"class":4451},[273,4539,4523],{"class":283},[273,4541,4542],{"class":1020},"fill",[273,4544,4475],{"class":283},[273,4546,4547],{"class":299},"'#email'",[273,4549,801],{"class":283},[273,4551,4552],{"class":299},"'user@test.com'",[273,4554,4481],{"class":283},[273,4556,4557,4559,4561,4564,4566,4569],{"class":275,"line":752},[273,4558,4520],{"class":4451},[273,4560,4523],{"class":283},[273,4562,4563],{"class":1020},"click",[273,4565,4475],{"class":283},[273,4567,4568],{"class":299},"'#login'",[273,4570,4481],{"class":283},[273,4572,4573,4575,4577,4580,4582,4585,4587,4590],{"class":275,"line":763},[273,4574,4520],{"class":4451},[273,4576,4523],{"class":283},[273,4578,4579],{"class":1020},"dragAndDrop",[273,4581,4475],{"class":283},[273,4583,4584],{"class":299},"'.task-item'",[273,4586,801],{"class":283},[273,4588,4589],{"class":299},"'.target-board'",[273,4591,4481],{"class":283},[273,4593,4594,4596,4599,4602,4605,4607,4610,4613,4616],{"class":275,"line":771},[273,4595,4520],{"class":4451},[273,4597,4598],{"class":1020}," expect",[273,4600,4601],{"class":283},"(page.",[273,4603,4604],{"class":1020},"locator",[273,4606,4475],{"class":283},[273,4608,4609],{"class":299},"'.success-toast'",[273,4611,4612],{"class":283},")).",[273,4614,4615],{"class":1020},"toBeVisible",[273,4617,4618],{"class":283},"();\n",[273,4620,4621],{"class":275,"line":781},[273,4622,4623],{"class":4445},"  \u002F\u002F Flag failures: screenshots, console logs\n",[273,4625,4626],{"class":275,"line":791},[273,4627,4628],{"class":283},"});\n",[13,4630,4631],{},"Evaluator runs this, parses results: \"Drag-drop fails on Safari mobile—fix pointer events.\"",[13,4633,4634,4635],{},"Case study from Archon V3: On a 1M-line codebase, generator self-evals passed 90% flawed UIs (missed accessibility bugs). Separate Evaluator caught 70% more issues via rubric + Playwright, boosting deployable PRs to 3.5\u002Fday per engineer. Stripe's 1,300 weekly PRs rely on similar adversarial grading atop primitives.",[17,4636,4637],{},[20,4638,343],{"href":340,"ariaDescribedBy":4639,"dataFootnoteRef":8,"id":4640},[24],"user-content-fnref-6-5",[13,4642,4643],{},"Rubric example:",[2156,4645,4646,4659],{},[2159,4647,4648],{},[2162,4649,4650,4653,4656],{},[2165,4651,4652],{},"Axis",[2165,4654,4655],{},"Weight",[2165,4657,4658],{},"Criteria",[2178,4660,4661,4672,4683,4693],{},[2162,4662,4663,4666,4669],{},[2183,4664,4665],{},"Design",[2183,4667,4668],{},"25%",[2183,4670,4671],{},"Figma-level alignment, responsive",[2162,4673,4674,4677,4680],{},[2183,4675,4676],{},"Originality",[2183,4678,4679],{},"20%",[2183,4681,4682],{},"Avoids cookie-cutter templates",[2162,4684,4685,4688,4690],{},[2183,4686,4687],{},"Craft",[2183,4689,4668],{},[2183,4691,4692],{},"Zero console errors, fast loads",[2162,4694,4695,4698,4701],{},[2183,4696,4697],{},"Functionality",[2183,4699,4700],{},"30%",[2183,4702,4703],{},"100% test pass, no regressions",[13,4705,4706],{},"Example score on a task manager generator output: Design 8\u002F10 (responsive but typography inconsistent), Originality 6\u002F10 (Kanban clone, lacks swipe gestures), Craft 9\u002F10 (95 Lighthouse, no leaks), Functionality 7\u002F10 (drag works desktop, notifications silent on reconnect). Weighted: 7.6. Fixes: \"Update fonts to Inter, add mobile swipes for originality, wire WebSocket reconnect.\"",[13,4708,4709],{},"Iterate until green. Production gold.",[46,4711,4713],{"id":4712},"_12-primitives-for-production-ready-agents","12 Primitives for Production-Ready Agents",[13,4715,4716,4717],{},"Production agents rely on 12 primitives from Anthropic's Claude Code leak—tool registries with 200+ options, tiered permissions, state persistence, token budgeting, observability, and more—to manage crashes, costs, real-world actions safely, and scale demos into reliable systems like Stripe's 1,300 weekly PRs with zero human code.",[17,4718,4719],{},[20,4720,34],{"href":31,"ariaDescribedBy":4721,"dataFootnoteRef":8,"id":127},[24],[13,4723,4724,4725,4730],{},"Nate B. Jones nails it: \"No permissions layer? It's a demo, not a product.\"",[17,4726,4727],{},[20,4728,34],{"href":31,"ariaDescribedBy":4729,"dataFootnoteRef":8,"id":162},[24]," Here's the list:",[183,4732,4733,4739,4745,4751,4757,4762,4768,4773,4779,4785,4791,4796],{},[186,4734,4735,4738],{},[636,4736,4737],{},"Tool registries",": 200+ tools, dynamic pools per session.",[186,4740,4741,4744],{},[636,4742,4743],{},"Tiered permissions",": Built-in (read), plugins (write), user (approve).",[186,4746,4747,4750],{},[636,4748,4749],{},"State persistence",": JSON sessions for messages\u002Ftokens\u002Fconfig.",[186,4752,4753,4756],{},[636,4754,4755],{},"Token budgeting",": Project usage, halt at limits.",[186,4758,4759,4761],{},[636,4760,1159],{},": Real-time events, no black boxes.",[186,4763,4764,4767],{},[636,4765,4766],{},"System logging",": Audit every action.",[186,4769,4770,4772],{},[636,4771,1162],{},": Harness tests pre-deploy.",[186,4774,4775,4778],{},[636,4776,4777],{},"Crash recovery",": Resume from last state.",[186,4780,4781,4784],{},[636,4782,4783],{},"Agent types",": Explore\u002Fplan\u002Fverify\u002Fguide\u002Fgeneral\u002Fstatus.",[186,4786,4787,4790],{},[636,4788,4789],{},"Workflow separation",": Agent state vs. task state.",[186,4792,4793,4795],{},[636,4794,1298],{},": Metrics dashboards.",[186,4797,4798,4801],{},[636,4799,4800],{},"Hooks",": Pre\u002Fpost-tool correction.",[13,4803,4804,4805,4811],{},"Stripe ships 1,300 PRs\u002Fweek zero human code using these.",[17,4806,4807],{},[20,4808,343],{"href":340,"ariaDescribedBy":4809,"dataFootnoteRef":8,"id":4810},[24],"user-content-fnref-6-6"," Skip them? Costs explode, bugs slip through.",[46,4813,4815],{"id":4814},"building-your-own-3-core-agent-harness","Building Your Own 3-Core-Agent Harness",[13,4817,4818,4819],{},"Build a 3-core-agent harness by defining YAML workflows in Archon V3 or SpecKit for DAG isolation of Planner, Generator, and Evaluator; connect via Claude teams or GSD-enhanced evaluation; incorporate primitives like persistence and permissions to deploy production systems in days, bypassing full framework overhead.",[17,4820,4821],{},[20,4822,343],{"href":340,"ariaDescribedBy":4823,"dataFootnoteRef":8,"id":4824},[24],"user-content-fnref-6-7",[13,4826,1478],{},[401,4828,4829,4835,4838,4841],{},[186,4830,4831,4832,987],{},"Define YAML nodes: ",[270,4833,4834],{},"planner: {model: claude-opus, role: PM}",[186,4836,4837],{},"Git worktrees for parallelism: One per agent run.",[186,4839,4840],{},"Hooks: Post-gen eval trigger.",[186,4842,4843],{},"Triggers: CLI\u002FSlack, multi-provider (Anthropic\u002FOpenAI).",[13,4845,4846],{},"Basic Python skeleton:",[264,4848,4850],{"className":1266,"code":4849,"language":1268,"meta":8,"style":8},"import yaml\nfrom anthropic import Anthropic\n\nworkflow = yaml.safe_load(open(\"harness.yaml\"))\nclient = Anthropic()\n\ndef run_harness(spec):\n    plan = client.messages.create(model=\"claude-opus-4.6\", role=\"planner\", ...)\n    gen = client.messages.create(model=\"claude-opus-4.6\", role=\"generator\", plan)\n    while eval_score(gen) \u003C 8:\n        gen = client.messages.create(..., previous=gen + fixes)\n    return gen\n",[270,4851,4852,4857,4862,4866,4871,4876,4880,4885,4890,4895,4900,4905],{"__ignoreMap":8},[273,4853,4854],{"class":275,"line":276},[273,4855,4856],{},"import yaml\n",[273,4858,4859],{"class":275,"line":287},[273,4860,4861],{},"from anthropic import Anthropic\n",[273,4863,4864],{"class":275,"line":303},[273,4865,1971],{"emptyLinePlaceholder":616},[273,4867,4868],{"class":275,"line":314},[273,4869,4870],{},"workflow = yaml.safe_load(open(\"harness.yaml\"))\n",[273,4872,4873],{"class":275,"line":724},[273,4874,4875],{},"client = Anthropic()\n",[273,4877,4878],{"class":275,"line":739},[273,4879,1971],{"emptyLinePlaceholder":616},[273,4881,4882],{"class":275,"line":752},[273,4883,4884],{},"def run_harness(spec):\n",[273,4886,4887],{"class":275,"line":763},[273,4888,4889],{},"    plan = client.messages.create(model=\"claude-opus-4.6\", role=\"planner\", ...)\n",[273,4891,4892],{"class":275,"line":771},[273,4893,4894],{},"    gen = client.messages.create(model=\"claude-opus-4.6\", role=\"generator\", plan)\n",[273,4896,4897],{"class":275,"line":781},[273,4898,4899],{},"    while eval_score(gen) \u003C 8:\n",[273,4901,4902],{"class":275,"line":791},[273,4903,4904],{},"        gen = client.messages.create(..., previous=gen + fixes)\n",[273,4906,4907],{"class":275,"line":809},[273,4908,4909],{},"    return gen\n",[13,4911,4912,4913,4920],{},"SpecKit style: Specify\u002Fplan\u002Ftasks\u002Fimplement.",[17,4914,4915],{},[20,4916,457],{"href":4917,"ariaDescribedBy":4918,"dataFootnoteRef":8,"id":4919},"#user-content-fn-7",[24],"user-content-fnref-7"," Scale to Stripe levels in weeks.",[46,4922,4924],{"id":4923},"trade-offs-misconceptions-and-when-to-use-frameworks","Trade-offs, Misconceptions, and When to Use Frameworks",[13,4926,4927,4930],{},[636,4928,4929],{},"3-core-agent harnesses"," cut complexity for strong LLMs but cost more than chains\u002Frouting; misconceptions include needing sub-agents (rare) or self-eval (flawed)—use frameworks only for weak models or specialized PRDs; audit via metrics like success rate\u002Fcost.",[13,4932,4933,4934,4939,4940,4946],{},"Agents flex dynamically but burn tokens—chains win for fixed paths.",[17,4935,4936],{},[20,4937,108],{"href":105,"ariaDescribedBy":4938,"dataFootnoteRef":8,"id":1474},[24]," Misconception: Sub-agents for everything; 90% tasks need three.",[17,4941,4942],{},[20,4943,26],{"href":22,"ariaDescribedBy":4944,"dataFootnoteRef":8,"id":4945},[24],"user-content-fnref-1-15"," Self-eval? Overconfident trash.",[13,4948,4949],{},"Frameworks fit smaller LLMs or edge PRDs. Audit: Track success\u002Fcost\u002Flatency pre\u002Fpost-strip.",[2156,4951,4952,4968],{},[2159,4953,4954],{},[2162,4955,4956,4959,4962,4965],{},[2165,4957,4958],{},"Setup",[2165,4960,4961],{},"Success Rate",[2165,4963,4964],{},"Cost",[2165,4966,4967],{},"Use When",[2178,4969,4970,4984,4998],{},[2162,4971,4972,4975,4978,4981],{},[2183,4973,4974],{},"Harness",[2183,4976,4977],{},"85%+",[2183,4979,4980],{},"Higher",[2183,4982,4983],{},"Strong LLMs, complex tasks",[2162,4985,4986,4989,4992,4995],{},[2183,4987,4988],{},"Framework",[2183,4990,4991],{},"70%",[2183,4993,4994],{},"Similar",[2183,4996,4997],{},"Weak models, sharding needed",[2162,4999,5000,5003,5006,5009],{},[2183,5001,5002],{},"Chains",[2183,5004,5005],{},"95%",[2183,5007,5008],{},"Lowest",[2183,5010,5011],{},"Predictable flows",[13,5013,5014],{},"Master metrics first.",[13,5016,5017,5018],{},"Pick a Git repo today. Strip your agent stack to Planner\u002FGenerator\u002FEvaluator using Archon V3 YAML. Run a full feature PR end-to-end. Score on the 4-axis rubric and compare success rates—watch reliability jump.",[17,5019,5020],{},[20,5021,343],{"href":340,"ariaDescribedBy":5022,"dataFootnoteRef":8,"id":5023},[24],"user-content-fnref-6-8",[391,5025,5027,5030],{"className":5026,"dataFootnotes":8},[394],[46,5028,399],{"className":5029,"id":24},[398],[401,5031,5032,5124,5140,5146,5162,5168,5219],{},[186,5033,5034,5035,414,5038,414,5043,414,5048,414,5053,414,5058,414,5063,414,5068,414,5073,414,5080,414,5087,414,5094,414,5101,414,5108,414,5116],{"id":405},"AI LABS summary of Anthropic experiments on agent frameworks. ",[20,5036,413],{"href":409,"ariaLabel":410,"className":5037,"dataFootnoteBackref":8},[412],[20,5039,413,5041],{"href":417,"ariaLabel":418,"className":5040,"dataFootnoteBackref":8},[412],[17,5042,34],{},[20,5044,413,5046],{"href":424,"ariaLabel":425,"className":5045,"dataFootnoteBackref":8},[412],[17,5047,79],{},[20,5049,413,5051],{"href":431,"ariaLabel":432,"className":5050,"dataFootnoteBackref":8},[412],[17,5052,108],{},[20,5054,413,5056],{"href":438,"ariaLabel":439,"className":5055,"dataFootnoteBackref":8},[412],[17,5057,146],{},[20,5059,413,5061],{"href":445,"ariaLabel":446,"className":5060,"dataFootnoteBackref":8},[412],[17,5062,343],{},[20,5064,413,5066],{"href":452,"ariaLabel":453,"className":5065,"dataFootnoteBackref":8},[412],[17,5067,457],{},[20,5069,413,5071],{"href":1628,"ariaLabel":1629,"className":5070,"dataFootnoteBackref":8},[412],[17,5072,1633],{},[20,5074,413,5078],{"href":5075,"ariaLabel":5076,"className":5077,"dataFootnoteBackref":8},"#user-content-fnref-1-9","Back to reference 1-9",[412],[17,5079,1686],{},[20,5081,413,5085],{"href":5082,"ariaLabel":5083,"className":5084,"dataFootnoteBackref":8},"#user-content-fnref-1-10","Back to reference 1-10",[412],[17,5086,1694],{},[20,5088,413,5092],{"href":5089,"ariaLabel":5090,"className":5091,"dataFootnoteBackref":8},"#user-content-fnref-1-11","Back to reference 1-11",[412],[17,5093,2263],{},[20,5095,413,5099],{"href":5096,"ariaLabel":5097,"className":5098,"dataFootnoteBackref":8},"#user-content-fnref-1-12","Back to reference 1-12",[412],[17,5100,2283],{},[20,5102,413,5106],{"href":5103,"ariaLabel":5104,"className":5105,"dataFootnoteBackref":8},"#user-content-fnref-1-13","Back to reference 1-13",[412],[17,5107,2322],{},[20,5109,413,5113],{"href":5110,"ariaLabel":5111,"className":5112,"dataFootnoteBackref":8},"#user-content-fnref-1-14","Back to reference 1-14",[412],[17,5114,5115],{},"14",[20,5117,413,5121],{"href":5118,"ariaLabel":5119,"className":5120,"dataFootnoteBackref":8},"#user-content-fnref-1-15","Back to reference 1-15",[412],[17,5122,5123],{},"15",[186,5125,5126,5127,414,5130,414,5135],{"id":460},"Claude Code Leak analysis by Nate B. Jones, AI News & Strategy Daily. ",[20,5128,413],{"href":464,"ariaLabel":465,"className":5129,"dataFootnoteBackref":8},[412],[20,5131,413,5133],{"href":469,"ariaLabel":470,"className":5132,"dataFootnoteBackref":8},[412],[17,5134,34],{},[20,5136,413,5138],{"href":476,"ariaLabel":477,"className":5137,"dataFootnoteBackref":8},[412],[17,5139,79],{},[186,5141,5142,5143],{"id":504},"Anthropic's 3 agents prior coverage. ",[20,5144,413],{"href":508,"ariaLabel":509,"className":5145,"dataFootnoteBackref":8},[412],[186,5147,5148,5149,414,5152,414,5157],{"id":534},"Agent Blueprint by Lukas Margerie. ",[20,5150,413],{"href":538,"ariaLabel":539,"className":5151,"dataFootnoteBackref":8},[412],[20,5153,413,5155],{"href":543,"ariaLabel":544,"className":5154,"dataFootnoteBackref":8},[412],[17,5156,34],{},[20,5158,413,5160],{"href":1743,"ariaLabel":1744,"className":5159,"dataFootnoteBackref":8},[412],[17,5161,79],{},[186,5163,5164,5165],{"id":550},"BMAD\u002FSuperpowers for PRD tasks. ",[20,5166,413],{"href":554,"ariaLabel":555,"className":5167,"dataFootnoteBackref":8},[412],[186,5169,5170,5171,414,5174,414,5179,414,5184,414,5191,414,5198,414,5205,414,5212],{"id":566},"Archon V3 by DIY Smart Code. ",[20,5172,413],{"href":570,"ariaLabel":571,"className":5173,"dataFootnoteBackref":8},[412],[20,5175,413,5177],{"href":575,"ariaLabel":576,"className":5176,"dataFootnoteBackref":8},[412],[17,5178,34],{},[20,5180,413,5182],{"href":582,"ariaLabel":583,"className":5181,"dataFootnoteBackref":8},[412],[17,5183,79],{},[20,5185,413,5189],{"href":5186,"ariaLabel":5187,"className":5188,"dataFootnoteBackref":8},"#user-content-fnref-6-4","Back to reference 6-4",[412],[17,5190,108],{},[20,5192,413,5196],{"href":5193,"ariaLabel":5194,"className":5195,"dataFootnoteBackref":8},"#user-content-fnref-6-5","Back to reference 6-5",[412],[17,5197,146],{},[20,5199,413,5203],{"href":5200,"ariaLabel":5201,"className":5202,"dataFootnoteBackref":8},"#user-content-fnref-6-6","Back to reference 6-6",[412],[17,5204,343],{},[20,5206,413,5210],{"href":5207,"ariaLabel":5208,"className":5209,"dataFootnoteBackref":8},"#user-content-fnref-6-7","Back to reference 6-7",[412],[17,5211,457],{},[20,5213,413,5217],{"href":5214,"ariaLabel":5215,"className":5216,"dataFootnoteBackref":8},"#user-content-fnref-6-8","Back to reference 6-8",[412],[17,5218,1633],{},[186,5220,5222,5223],{"id":5221},"user-content-fn-7","GitHub SpecKit spec-centric agents. ",[20,5224,413],{"href":5225,"ariaLabel":5226,"className":5227,"dataFootnoteBackref":8},"#user-content-fnref-7","Back to reference 7",[412],[588,5229,5230],{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":8,"searchDepth":287,"depth":287,"links":5232},[5233,5234,5235,5236,5237,5238,5239,5240,5241],{"id":4173,"depth":287,"text":4174},{"id":1860,"depth":287,"text":4258},{"id":4293,"depth":287,"text":4294},{"id":4339,"depth":287,"text":4340},{"id":4376,"depth":287,"text":4377},{"id":4712,"depth":287,"text":4713},{"id":4814,"depth":287,"text":4815},{"id":4923,"depth":287,"text":4924},{"id":24,"depth":287,"text":399},"Production agent systems thrive with a 3-core-agent harness—Planner for high-level specs, Generator for implementation, Evaluator for rigorous checks. Ditch bloated frameworks. Modern LLMs like Claude Opus 4.6 use 1M-token contexts and coherence for reliable, scalable outputs on complex tasks.1",[3371,5244,1782,5245,1783,5246,5247,5248],"production ai agents","claude opus agents","anthropic agent experiments","ai harness engineering","agent primitives",{},"Ditch bloated frameworks like BMAD and GSD. Build production agent systems with a 3-core-agent harness: Planner, Generator, Evaluator. Leverage Claude Opus for reliable, scalable AI outputs without error cascades.","3-Core-Agent Harness for Production AI Agents","\u002Farticles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-e1-budget-8k-article",[],{"why-agent-frameworks-fail-production-systems":4174,"what-is-the-3-core-agent-harness":4258,"the-planner-defining-high-level-objectives":4294,"the-generator-implementation-without-micro-management":4340,"the-evaluator-adversarial-checks-with-graded-rubrics":4377,"12-primitives-for-production-ready-agents":4713,"building-your-own-3-core-agent-harness":4815,"trade-offs-misconceptions-and-when-to-use-frameworks":4924},{"title":4147,"description":5242},{"loc":5252},"the-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks","articles\u002Fthe-3-core-agent-harness-why-production-agent-systems-need-planner-generator-evaluator-not-frameworks-e1-budget-8k-article","m2bi_KTVXWqB41yv7BbLfXKhYm8AEsWwCyKDcNTfZt4",[5261,5264,5266,5269,5271,5274,5277,5280,5283,5285,5287,5289,5291,5293,5295,5297,5300,5302,5304,5306,5308,5310,5312,5315,5317,5319,5321,5323,5325,5327,5329,5331,5333,5335,5337,5339,5341,5343,5345,5347,5349,5351,5353,5355,5357,5360,5362,5364,5366,5368,5370,5372,5374,5376,5378,5380,5382,5384,5386,5388,5390,5392,5394,5396,5398,5400,5402,5404,5406,5408,5410,5412,5414,5416,5418,5420,5422,5424,5426,5428,5430,5432,5434,5436,5438,5440,5442,5444,5446,5448,5450,5452,5454,5456,5458,5460,5462,5464,5466,5468,5470,5472,5474,5476,5478,5480,5482,5484,5486,5488,5490,5492,5494,5496,5498,5500,5502,5504,5506,5508,5510,5512,5514,5516,5518,5520,5522,5524,5526,5528,5530,5532,5534,5536,5538,5540,5542,5544,5546,5548,5550,5552,5554,5556,5558,5560,5562,5564,5566,5568,5570,5572,5574,5576,5578,5580,5582,5584,5586,5588,5590,5592,5594,5596,5598,5600,5602,5604,5606,5608,5610,5612,5614,5616,5618,5620,5622,5625,5627,5629,5631,5633,5635,5637,5639,5641,5643,5645,5647,5649,5651,5653,5655,5657,5659,5661,5663,5665,5667,5669,5671,5673,5675,5677,5679,5681,5683,5685,5687,5689,5691,5693,5695,5697,5699,5701,5703,5705,5707,5709,5712,5714,5716,5718,5720,5722,5724,5726,5728,5730,5732,5734,5736,5738,5740,5742,5744,5746,5748,5750,5752,5754,5756,5758,5760,5762,5764,5766,5768,5770,5772,5774,5776,5778,5780,5782,5784,5786,5788,5790,5792,5794,5796,5798,5800,5802,5804,5806,5808,5810,5812,5814,5816,5818,5820,5822,5824,5826,5828,5830,5832,5834,5836,5838,5840,5842,5844,5846,5848,5850,5852,5854,5856,5858,5860,5862,5864,5866,5868,5870,5872,5874,5876,5878,5880,5882,5884,5886,5888,5890,5892,5894,5896,5898,5900,5902,5904,5906,5908,5910,5912,5914,5916,5918,5920,5922,5924,5926,5928,5930,5932,5934,5936,5938,5940,5942,5944,5946,5948,5950,5952,5954,5956,5958,5960,5962,5964,5966,5968,5970,5972,5974,5976,5978,5980,5982,5984,5986,5988,5990,5992,5994,5996,5998,6000,6002,6004,6006,6008,6010,6012,6014,6016,6018,6020,6022,6024,6026,6028,6030,6032,6034,6036,6038,6040,6042,6044,6046,6048,6050,6052,6054,6056,6058,6060,6062,6064,6066,6068,6070,6072,6074,6076,6078,6080,6082,6084,6086,6088,6090,6092,6094,6096,6098,6100,6102,6104,6106,6108,6110,6112,6114,6116,6118,6120,6122,6124,6126,6128,6130,6132,6134,6136,6138,6140,6142,6144,6146,6148,6150,6152,6154,6156,6158,6160,6162,6164,6166,6168,6170,6172,6174,6176,6178,6180,6182,6184,6186,6188,6190,6192,6194,6196,6198,6200,6202,6204,6206,6208,6210,6212,6214,6216,6218,6220,6222,6224,6226,6228,6230,6232,6234,6236,6238,6240,6242,6244,6246,6248,6250,6252,6254,6256,6258,6260,6262,6264,6266,6268,6270,6272,6274,6276,6278,6280,6282,6284,6286,6288,6290,6292,6294,6296,6298,6300,6302,6304,6306,6308,6310,6312,6314,6316,6318,6320,6322,6324,6326,6328,6330,6332,6334,6336,6338,6340,6342,6344,6346,6348,6350,6352,6354,6356,6358,6360,6362,6364,6366,6368,6370,6372,6374,6376,6378,6380,6382,6384,6386,6388,6390,6392,6394,6396,6398,6400,6402,6404,6406,6408,6410,6412,6414,6416,6418,6420,6422,6424,6426,6428,6430,6432,6434,6436,6438,6440,6442,6444,6446,6448,6450,6452,6454,6456,6458,6460,6462,6464,6466,6468,6470,6472,6474,6476,6478,6480,6482,6484,6486,6488,6490,6492,6494,6496,6498,6500,6502,6504,6506,6508,6510,6512,6514,6516,6518,6520,6522,6524,6526,6528,6530,6532,6534,6536,6538,6540,6542,6544,6546,6548,6550,6552,6554,6556,6558,6560,6562,6564,6566,6568,6570,6572,6574,6576,6578,6580,6582,6584,6586,6588,6590,6592,6594,6596,6598,6600,6602,6604,6606,6608,6610,6612,6614,6616,6618,6620,6622,6624,6626,6628,6630,6632,6634,6636,6638,6640,6642,6644,6646,6648,6650,6652,6654,6656,6658,6660,6662,6664,6666,6668,6670,6672,6674,6676,6678,6680,6682,6684,6686,6688,6690,6692,6694,6696,6698,6700,6702,6704,6706,6708,6710,6712,6714,6716,6718,6720,6722,6724,6726,6728,6730,6732,6734,6736,6738,6740,6742,6744,6746,6748,6750,6752,6754,6756,6758,6760,6762,6764,6766,6768,6770,6772,6774,6776,6778,6780,6782,6784,6786,6788,6790,6792,6794,6796,6798,6800,6802,6804,6806,6808,6810,6812,6814,6816,6818,6820,6822,6824,6826,6828,6830,6832,6834,6836,6838,6840,6842,6844,6846,6848,6850,6852,6854,6856,6858,6860,6862,6864,6866,6868,6870,6872,6874,6876,6878,6880,6882,6884,6886,6888,6890,6892,6894,6896,6898,6900,6902,6904,6906,6908,6910,6912,6914,6916,6918,6920,6922,6924,6926,6928,6930,6932,6934,6936,6938,6940,6942,6944,6946,6948,6950,6952,6954,6956,6958,6960,6962,6964,6966,6968,6970,6972,6974,6976,6978,6980,6982,6984,6986,6988,6990,6992,6994,6996,6998,7000,7002,7004,7006,7008,7010,7012,7014,7016,7018,7020,7022,7024,7026,7028,7030,7032,7034,7036,7038,7040,7042,7044,7046,7048,7050,7052,7054,7056,7058,7060,7062,7064,7066,7068,7070,7072,7074,7076,7078,7080,7082,7084,7086,7088,7090,7092,7094,7096,7098,7100,7102,7104,7106,7108,7110,7112,7114,7116,7118,7120,7122,7124,7126,7128,7130,7132,7134,7136,7138,7140,7142,7144,7146,7148,7150,7152,7154,7156,7158,7160,7162,7164,7166,7168,7170,7172,7174,7176,7178,7180,7182,7184,7186,7188,7190,7192,7194,7196,7198,7200,7202,7204,7206,7208,7210,7212,7214,7216,7218,7220,7222,7224,7226,7228,7230,7232,7234,7236,7238,7240,7242,7244,7246,7248,7250,7252,7254,7256,7258,7260,7262,7264,7266,7268,7270,7272,7274,7276,7278,7280,7282,7284,7286,7288,7290,7292,7294,7296,7298,7300,7302,7304,7306,7308,7310,7312,7314,7316,7318,7320,7322,7324,7326,7328,7330,7332,7334,7336,7338,7340,7342,7344,7346,7348,7350,7352,7354,7356,7358,7360,7362,7364,7366,7368,7370,7372,7374,7376,7378,7380,7382,7384,7386,7388,7390,7392,7394,7396,7398,7400,7402,7404,7406,7408,7410,7412,7414,7416,7418,7420,7422,7424,7426,7428,7430,7432,7434,7436,7438,7440,7442,7444,7446,7448,7450,7452,7454,7456,7458,7460,7462,7464,7466,7468,7470,7472,7474,7476,7478,7480,7482,7484,7486,7488,7490,7492,7494,7496,7498,7500,7502,7504,7506,7508,7510,7512,7514,7516,7518,7520,7522,7524,7526,7528,7530,7532,7534,7536,7538,7540,7542,7544,7546,7548,7550,7552,7554,7556,7558,7560,7562,7564,7566,7568,7570,7572,7574,7576,7578,7580,7582,7584,7586,7588,7590,7592,7594,7596,7598,7600,7602,7604,7606,7608,7610,7612,7614,7616,7618,7620,7622,7624,7626,7628,7630,7632,7634,7636,7638,7640,7642,7644,7646,7648,7650,7652,7654,7656,7658,7660,7662,7664,7666,7668,7670,7672,7674,7676,7678,7680,7682,7684,7686,7688,7690,7692,7694,7696,7698,7700,7702,7704,7706,7708,7710,7712,7714,7716,7718,7720,7722,7724,7726,7728,7730,7732,7734,7736,7738,7740,7742,7744,7746,7748,7750,7752,7754,7756,7758,7760,7762,7764,7766,7768,7770,7772,7774,7776,7778,7780,7782,7784,7786,7788,7790,7792,7794,7796,7798,7800,7802,7804,7806,7808,7810,7812,7814,7816,7818,7820,7822,7824,7826,7828,7830,7832,7834,7836,7838,7840,7842,7844,7846,7848,7850,7852,7854,7856,7858,7860,7862,7864,7866,7868,7870,7872,7874,7876,7878,7880,7882,7884,7886,7888,7890,7892,7894,7896,7898,7900,7902,7904,7906,7908,7910,7912,7914,7916,7918,7920,7922,7924,7926,7928,7930,7932,7934,7936,7938,7940,7942,7944,7946,7948,7950,7952,7954,7956,7958,7960,7962,7964,7966,7968,7970,7972,7974,7976,7978,7980,7982,7984,7986,7988,7990,7992,7994,7996,7998,8000,8002,8004,8006,8008,8010,8012,8014,8016,8018,8020,8022,8024,8026,8028,8030,8032,8034,8036,8038,8040,8042,8044,8046,8048,8050,8052,8054,8056,8058,8060,8062,8064,8066,8068,8070,8072,8074,8076,8078,8080,8082,8084,8086,8088,8090,8092,8094,8096,8098,8100,8102,8104,8106,8108,8110,8112,8114,8116,8118,8120,8122,8124,8126,8128,8130,8132,8134,8136,8138,8140,8142,8144,8146,8148,8150,8152,8154,8156,8158,8160,8162,8164,8166,8168,8170,8172,8174,8176,8178,8180,8182,8184,8186,8188,8190,8192,8194,8196,8198,8200,8202,8204,8206,8208,8210,8212,8214,8216,8218,8220,8222,8224,8226,8228,8230,8232,8234,8236,8238,8240,8242,8244,8246,8248,8250,8252,8254,8256,8258,8260,8262,8264,8266,8268,8270,8272,8274,8276,8278,8280,8282,8284,8286,8288,8290,8292,8294,8296,8298,8300,8302,8304,8306,8308,8310,8312,8314,8316,8318,8320,8322,8324,8326,8328,8330,8332,8334,8336,8338,8340,8342,8344,8346,8348,8350,8352,8354,8356,8358,8360,8362,8364,8366,8368,8370,8372,8374,8376,8378,8380,8382,8384,8386,8388,8390,8392,8394,8396,8398,8400,8402,8404,8406,8408,8410,8412,8414,8416,8418,8420,8422,8424,8426,8428,8430,8432,8434,8436,8438,8440,8442,8444,8446,8448,8450,8452,8454,8456,8458,8460,8462,8464,8466,8468,8470,8472,8474,8476,8478,8480,8482,8484,8486,8488,8490,8492,8494,8496,8498,8500,8502,8504,8506,8508,8510,8512,8514,8516,8518,8520,8522,8524,8526,8528,8530,8532,8534,8536,8538,8540,8542,8544,8546,8548,8550,8552,8554,8556,8558,8560,8562,8564,8566,8568,8570,8572,8574,8576,8578,8580,8582,8584,8586,8588,8590,8592,8594,8596,8598,8600,8602,8604,8606,8608,8610,8612,8614,8616,8618,8620,8622,8624,8626,8628,8630,8632,8634,8636,8638,8640,8642,8644,8646,8648,8650,8652,8654,8656,8658,8660,8662,8664,8666,8668,8670,8672,8674,8676,8678,8680,8682,8684,8686,8688,8690,8692,8694,8696,8698,8700,8702,8704,8706,8708,8710,8712,8714,8716,8718,8720,8722,8724,8726,8728,8730,8732,8734,8736,8738,8740,8742,8744,8746,8748,8750,8752,8754,8756,8758,8760,8762,8764,8766,8768,8770,8772,8774,8776,8778,8780,8782,8784,8786,8788,8790,8792,8794,8796,8798,8800,8802,8804,8806,8808,8810,8812,8814,8816,8818,8820,8822,8824,8826,8828,8830,8832,8834,8836,8838,8840,8842,8844,8846,8848,8850,8852,8854,8856,8858,8860,8862,8864,8866,8868,8870,8872,8874,8876,8878,8880,8882,8884,8886,8888,8890,8892,8894,8896,8898,8900,8902,8904,8906,8908,8910,8912,8914,8916,8918,8920,8922,8924,8926,8928,8930,8932,8934,8936,8938,8940,8942,8944,8946,8948,8950,8952,8954,8956,8958,8960,8962,8964,8966,8968,8970,8972,8974,8976,8978,8980,8982,8984,8986,8988,8990,8992,8994,8996,8998,9000,9002,9004,9006,9008,9010,9012,9014,9016,9018,9020,9022,9024,9026,9028,9030,9032,9034,9036,9038,9040,9042,9044,9046,9048,9050,9052,9054,9056,9058,9060,9062,9064,9066,9068,9070,9072,9074,9076,9078,9080,9082,9084,9086,9088,9090,9092,9094,9096,9098,9100,9102,9104,9106,9108,9110,9112,9114,9116,9118,9120,9122,9124,9126,9128,9130,9132,9134,9136,9138,9140,9142,9144,9146,9148,9150,9152,9154,9156,9158,9160,9162,9164,9166,9168,9170,9172,9174,9176,9178,9180,9182,9184,9186,9188,9190,9192,9194,9196,9198,9200,9202,9204,9206,9208,9210,9212,9214,9216,9218,9220,9222,9224,9226,9228,9230,9232,9234,9236,9238,9240,9242,9244,9246,9248,9250,9252,9254,9256,9258,9260,9262,9264,9266,9268,9270,9272,9274,9276,9278,9280,9282,9284,9286,9288,9290,9292,9294,9296,9298,9300,9302,9304,9306,9308,9310,9312,9314,9316,9318,9320,9322,9324,9326,9328,9330,9332,9334,9336,9338,9340,9342,9344,9346,9348,9350,9352,9354,9356,9358,9360,9362,9364,9366,9368,9370,9372,9374,9376,9378,9380,9382,9384,9386,9388,9390,9392,9394,9396,9398,9400,9402,9404,9406,9408,9410,9412,9414,9416,9418,9420,9422,9424,9426,9428,9430,9432,9434,9436,9438,9440,9442,9444,9446,9448,9450,9452,9454,9456,9458,9460,9462,9464,9466,9468,9470,9472,9474,9476,9478,9480,9482,9484,9486,9488,9490,9492,9494,9496,9498,9500,9502,9504,9506,9508,9510,9512,9514,9516,9518,9520,9522,9524,9526,9528,9530,9532,9534,9536,9538,9540,9542,9544,9546,9548,9550,9552,9554,9556,9558,9560,9562,9564,9566,9568,9570,9572,9574,9576,9578,9580,9582,9584,9586,9588,9590,9592,9594,9596,9598,9600,9602,9604,9606,9608,9610,9612,9614,9616,9618,9620,9622,9624,9626,9628,9630,9632,9634,9636,9638,9640,9642,9644,9646,9648,9650,9652,9654,9656,9658,9660,9662,9664],{"categories":5262},[5263],"Business & SaaS",{"categories":5265},[5263],{"categories":5267},[5268],"AI News & Trends",{"categories":5270},[],{"categories":5272},[5273],"AI Automation",{"categories":5275},[5276],"Marketing & Growth",{"categories":5278},[5279],"Design & Frontend",{"categories":5281},[5282],"Software Engineering",{"categories":5284},[5273],{"categories":5286},[],{"categories":5288},[5279],{"categories":5290},[5279],{"categories":5292},[5273],{"categories":5294},[5279],{"categories":5296},[5279],{"categories":5298},[5299],"AI & LLMs",{"categories":5301},[5279],{"categories":5303},[5279],{"categories":5305},[],{"categories":5307},[5279],{"categories":5309},[5279],{"categories":5311},[5299],{"categories":5313},[5314],"Developer Productivity",{"categories":5316},[5299],{"categories":5318},[5299],{"categories":5320},[5299],{"categories":5322},[5268],{"categories":5324},[5299],{"categories":5326},[5273],{"categories":5328},[5263],{"categories":5330},[5268],{"categories":5332},[5276],{"categories":5334},[],{"categories":5336},[],{"categories":5338},[5273],{"categories":5340},[5273],{"categories":5342},[5273],{"categories":5344},[5276],{"categories":5346},[5299],{"categories":5348},[5314],{"categories":5350},[5268],{"categories":5352},[],{"categories":5354},[],{"categories":5356},[],{"categories":5358},[5359],"Data Science & Visualization",{"categories":5361},[],{"categories":5363},[5273],{"categories":5365},[5282],{"categories":5367},[5273],{"categories":5369},[5273],{"categories":5371},[5299],{"categories":5373},[5276],{"categories":5375},[5273],{"categories":5377},[],{"categories":5379},[],{"categories":5381},[],{"categories":5383},[5279],{"categories":5385},[5279],{"categories":5387},[5273],{"categories":5389},[5276],{"categories":5391},[5314],{"categories":5393},[5279],{"categories":5395},[5299],{"categories":5397},[5282],{"categories":5399},[5299],{"categories":5401},[],{"categories":5403},[5273],{"categories":5405},[5299],{"categories":5407},[5314],{"categories":5409},[5314],{"categories":5411},[],{"categories":5413},[5276],{"categories":5415},[5263],{"categories":5417},[5299],{"categories":5419},[5263],{"categories":5421},[5263],{"categories":5423},[5273],{"categories":5425},[5276],{"categories":5427},[5273],{"categories":5429},[5263],{"categories":5431},[5273],{"categories":5433},[5279],{"categories":5435},[5299],{"categories":5437},[5279],{"categories":5439},[5299],{"categories":5441},[5263],{"categories":5443},[5299],{"categories":5445},[5276],{"categories":5447},[],{"categories":5449},[5299],{"categories":5451},[5263],{"categories":5453},[],{"categories":5455},[5268],{"categories":5457},[5282],{"categories":5459},[],{"categories":5461},[5299],{"categories":5463},[5279],{"categories":5465},[5299],{"categories":5467},[5279],{"categories":5469},[],{"categories":5471},[5273],{"categories":5473},[],{"categories":5475},[],{"categories":5477},[],{"categories":5479},[5299],{"categories":5481},[],{"categories":5483},[5299],{"categories":5485},[5299],{"categories":5487},[5279],{"categories":5489},[5299],{"categories":5491},[5314],{"categories":5493},[5273],{"categories":5495},[5276],{"categories":5497},[5314],{"categories":5499},[5314],{"categories":5501},[5314],{"categories":5503},[5276],{"categories":5505},[5276],{"categories":5507},[5299],{"categories":5509},[5299],{"categories":5511},[5279],{"categories":5513},[5263],{"categories":5515},[5279],{"categories":5517},[5282],{"categories":5519},[5263],{"categories":5521},[5263],{"categories":5523},[5263],{"categories":5525},[5279],{"categories":5527},[],{"categories":5529},[],{"categories":5531},[5299],{"categories":5533},[5299],{"categories":5535},[5282],{"categories":5537},[5299],{"categories":5539},[5299],{"categories":5541},[],{"categories":5543},[5299],{"categories":5545},[5299],{"categories":5547},[],{"categories":5549},[5299],{"categories":5551},[5268],{"categories":5553},[5268],{"categories":5555},[],{"categories":5557},[],{"categories":5559},[5276],{"categories":5561},[5276],{"categories":5563},[5282],{"categories":5565},[5299],{"categories":5567},[],{"categories":5569},[],{"categories":5571},[5273],{"categories":5573},[5299],{"categories":5575},[5299],{"categories":5577},[],{"categories":5579},[5299,5263],{"categories":5581},[5299],{"categories":5583},[],{"categories":5585},[5299],{"categories":5587},[5299],{"categories":5589},[],{"categories":5591},[],{"categories":5593},[5273],{"categories":5595},[5299],{"categories":5597},[5299],{"categories":5599},[5273],{"categories":5601},[5299],{"categories":5603},[],{"categories":5605},[],{"categories":5607},[5299],{"categories":5609},[],{"categories":5611},[5299],{"categories":5613},[5299],{"categories":5615},[],{"categories":5617},[5273],{"categories":5619},[5279],{"categories":5621},[],{"categories":5623},[5273,5624],"DevOps & Cloud",{"categories":5626},[5299],{"categories":5628},[5273],{"categories":5630},[5299],{"categories":5632},[],{"categories":5634},[],{"categories":5636},[],{"categories":5638},[],{"categories":5640},[5299],{"categories":5642},[5273],{"categories":5644},[],{"categories":5646},[5273],{"categories":5648},[],{"categories":5650},[5299],{"categories":5652},[],{"categories":5654},[],{"categories":5656},[],{"categories":5658},[],{"categories":5660},[5273],{"categories":5662},[5279],{"categories":5664},[5299],{"categories":5666},[5276],{"categories":5668},[5268],{"categories":5670},[5263],{"categories":5672},[5314],{"categories":5674},[],{"categories":5676},[5273],{"categories":5678},[5273],{"categories":5680},[5299],{"categories":5682},[],{"categories":5684},[],{"categories":5686},[],{"categories":5688},[5273],{"categories":5690},[],{"categories":5692},[5273],{"categories":5694},[5273],{"categories":5696},[5268],{"categories":5698},[5273],{"categories":5700},[5299],{"categories":5702},[],{"categories":5704},[5299],{"categories":5706},[],{"categories":5708},[5268],{"categories":5710},[5273,5711],"Product Strategy",{"categories":5713},[5282],{"categories":5715},[5624],{"categories":5717},[5711],{"categories":5719},[5299],{"categories":5721},[5273],{"categories":5723},[],{"categories":5725},[5268],{"categories":5727},[5268],{"categories":5729},[5273],{"categories":5731},[],{"categories":5733},[5273],{"categories":5735},[5299],{"categories":5737},[5299],{"categories":5739},[5314],{"categories":5741},[5299],{"categories":5743},[],{"categories":5745},[5299,5282],{"categories":5747},[5268],{"categories":5749},[5299],{"categories":5751},[5268],{"categories":5753},[5273],{"categories":5755},[5268],{"categories":5757},[],{"categories":5759},[5282],{"categories":5761},[5263],{"categories":5763},[],{"categories":5765},[5273],{"categories":5767},[5273],{"categories":5769},[5273],{"categories":5771},[5273],{"categories":5773},[5263],{"categories":5775},[5279],{"categories":5777},[5276],{"categories":5779},[],{"categories":5781},[5273],{"categories":5783},[],{"categories":5785},[5268],{"categories":5787},[5268],{"categories":5789},[5268],{"categories":5791},[5273],{"categories":5793},[5268],{"categories":5795},[5299],{"categories":5797},[5314],{"categories":5799},[5299],{"categories":5801},[5282],{"categories":5803},[5299,5314],{"categories":5805},[5314],{"categories":5807},[5314],{"categories":5809},[5314],{"categories":5811},[5314],{"categories":5813},[5299],{"categories":5815},[],{"categories":5817},[],{"categories":5819},[5276],{"categories":5821},[],{"categories":5823},[5299],{"categories":5825},[5314],{"categories":5827},[5299],{"categories":5829},[5279],{"categories":5831},[5282],{"categories":5833},[],{"categories":5835},[5299],{"categories":5837},[5314],{"categories":5839},[5276],{"categories":5841},[5268],{"categories":5843},[5282],{"categories":5845},[5299],{"categories":5847},[],{"categories":5849},[5282],{"categories":5851},[5279],{"categories":5853},[5263],{"categories":5855},[5263],{"categories":5857},[],{"categories":5859},[5279],{"categories":5861},[5263],{"categories":5863},[5268],{"categories":5865},[5314],{"categories":5867},[5273],{"categories":5869},[5273],{"categories":5871},[5299],{"categories":5873},[5299],{"categories":5875},[5268],{"categories":5877},[5268],{"categories":5879},[5314],{"categories":5881},[5268],{"categories":5883},[],{"categories":5885},[5711],{"categories":5887},[5273],{"categories":5889},[5268],{"categories":5891},[5268],{"categories":5893},[5268],{"categories":5895},[5299],{"categories":5897},[5273],{"categories":5899},[5273],{"categories":5901},[5263],{"categories":5903},[5263],{"categories":5905},[5299],{"categories":5907},[5268],{"categories":5909},[],{"categories":5911},[5299],{"categories":5913},[5263],{"categories":5915},[5273],{"categories":5917},[5273],{"categories":5919},[5273],{"categories":5921},[5279],{"categories":5923},[5273],{"categories":5925},[5314],{"categories":5927},[5268],{"categories":5929},[5268],{"categories":5931},[5268],{"categories":5933},[5268],{"categories":5935},[5268],{"categories":5937},[],{"categories":5939},[],{"categories":5941},[5314],{"categories":5943},[5268],{"categories":5945},[5268],{"categories":5947},[5268],{"categories":5949},[],{"categories":5951},[5299],{"categories":5953},[],{"categories":5955},[],{"categories":5957},[5279],{"categories":5959},[5263],{"categories":5961},[],{"categories":5963},[5268],{"categories":5965},[5273],{"categories":5967},[5273],{"categories":5969},[5273],{"categories":5971},[5276],{"categories":5973},[5273],{"categories":5975},[],{"categories":5977},[5268],{"categories":5979},[5268],{"categories":5981},[5299],{"categories":5983},[],{"categories":5985},[5276],{"categories":5987},[5276],{"categories":5989},[5299],{"categories":5991},[5268],{"categories":5993},[5263],{"categories":5995},[5282],{"categories":5997},[5299],{"categories":5999},[],{"categories":6001},[5299],{"categories":6003},[5299],{"categories":6005},[5282],{"categories":6007},[5299],{"categories":6009},[5299],{"categories":6011},[5299],{"categories":6013},[5276],{"categories":6015},[5268],{"categories":6017},[5299],{"categories":6019},[5299],{"categories":6021},[5268],{"categories":6023},[5273],{"categories":6025},[5314],{"categories":6027},[5263],{"categories":6029},[5299],{"categories":6031},[5314],{"categories":6033},[5314],{"categories":6035},[],{"categories":6037},[5276],{"categories":6039},[5268],{"categories":6041},[5268],{"categories":6043},[5314],{"categories":6045},[5273],{"categories":6047},[5273],{"categories":6049},[5273],{"categories":6051},[5273],{"categories":6053},[5279],{"categories":6055},[5299],{"categories":6057},[5299],{"categories":6059},[5711],{"categories":6061},[5299],{"categories":6063},[5299],{"categories":6065},[5273],{"categories":6067},[5263],{"categories":6069},[5276],{"categories":6071},[],{"categories":6073},[5263],{"categories":6075},[5263],{"categories":6077},[],{"categories":6079},[5279],{"categories":6081},[5299],{"categories":6083},[],{"categories":6085},[],{"categories":6087},[5268],{"categories":6089},[5268],{"categories":6091},[5268],{"categories":6093},[5268],{"categories":6095},[],{"categories":6097},[5268],{"categories":6099},[5299],{"categories":6101},[5299],{"categories":6103},[],{"categories":6105},[5268],{"categories":6107},[5268],{"categories":6109},[5263],{"categories":6111},[5299],{"categories":6113},[],{"categories":6115},[],{"categories":6117},[5268],{"categories":6119},[5268],{"categories":6121},[5268],{"categories":6123},[5299],{"categories":6125},[5268],{"categories":6127},[5268],{"categories":6129},[5268],{"categories":6131},[5268],{"categories":6133},[5268],{"categories":6135},[],{"categories":6137},[5273],{"categories":6139},[5299],{"categories":6141},[5276],{"categories":6143},[5263],{"categories":6145},[5273],{"categories":6147},[5299],{"categories":6149},[],{"categories":6151},[5276],{"categories":6153},[5268],{"categories":6155},[5268],{"categories":6157},[5268],{"categories":6159},[5268],{"categories":6161},[5314],{"categories":6163},[5282],{"categories":6165},[],{"categories":6167},[5299],{"categories":6169},[5273],{"categories":6171},[5273],{"categories":6173},[5273],{"categories":6175},[5624],{"categories":6177},[5273],{"categories":6179},[5299],{"categories":6181},[5299],{"categories":6183},[5282],{"categories":6185},[5624],{"categories":6187},[5359],{"categories":6189},[5299],{"categories":6191},[5359],{"categories":6193},[],{"categories":6195},[5276],{"categories":6197},[5276],{"categories":6199},[5279],{"categories":6201},[5624],{"categories":6203},[5273],{"categories":6205},[5299],{"categories":6207},[5299],{"categories":6209},[5273],{"categories":6211},[5273],{"categories":6213},[5273],{"categories":6215},[5314],{"categories":6217},[5314],{"categories":6219},[5273],{"categories":6221},[5273],{"categories":6223},[],{"categories":6225},[5273],{"categories":6227},[5273],{"categories":6229},[5299],{"categories":6231},[5359],{"categories":6233},[5273],{"categories":6235},[5273],{"categories":6237},[5273],{"categories":6239},[5273],{"categories":6241},[5263],{"categories":6243},[5279],{"categories":6245},[5268],{"categories":6247},[5282],{"categories":6249},[5624],{"categories":6251},[5282],{"categories":6253},[5359],{"categories":6255},[],{"categories":6257},[5282],{"categories":6259},[],{"categories":6261},[],{"categories":6263},[5282],{"categories":6265},[5299],{"categories":6267},[],{"categories":6269},[],{"categories":6271},[],{"categories":6273},[5263],{"categories":6275},[],{"categories":6277},[],{"categories":6279},[5359],{"categories":6281},[5299],{"categories":6283},[5624],{"categories":6285},[5299],{"categories":6287},[],{"categories":6289},[5273],{"categories":6291},[5314],{"categories":6293},[5314],{"categories":6295},[5276],{"categories":6297},[5276],{"categories":6299},[5276],{"categories":6301},[5624],{"categories":6303},[5282],{"categories":6305},[5273],{"categories":6307},[5263],{"categories":6309},[5263],{"categories":6311},[5282],{"categories":6313},[5279],{"categories":6315},[5359],{"categories":6317},[5279],{"categories":6319},[],{"categories":6321},[5299],{"categories":6323},[5273],{"categories":6325},[5273],{"categories":6327},[5314],{"categories":6329},[5273],{"categories":6331},[5273],{"categories":6333},[5279],{"categories":6335},[5279],{"categories":6337},[5273],{"categories":6339},[5624],{"categories":6341},[5299],{"categories":6343},[],{"categories":6345},[5276],{"categories":6347},[5273],{"categories":6349},[5263],{"categories":6351},[5273],{"categories":6353},[5273],{"categories":6355},[],{"categories":6357},[5299],{"categories":6359},[5273],{"categories":6361},[5273],{"categories":6363},[5314],{"categories":6365},[5273],{"categories":6367},[5299],{"categories":6369},[],{"categories":6371},[5273],{"categories":6373},[],{"categories":6375},[5279],{"categories":6377},[5314],{"categories":6379},[5299],{"categories":6381},[5282],{"categories":6383},[5279],{"categories":6385},[5314],{"categories":6387},[5359],{"categories":6389},[5314],{"categories":6391},[],{"categories":6393},[5299],{"categories":6395},[5299],{"categories":6397},[5711],{"categories":6399},[5282],{"categories":6401},[5299,5273],{"categories":6403},[5273],{"categories":6405},[5299],{"categories":6407},[5273],{"categories":6409},[5273,5282],{"categories":6411},[5273],{"categories":6413},[5299],{"categories":6415},[],{"categories":6417},[5314],{"categories":6419},[5299],{"categories":6421},[5273],{"categories":6423},[5299],{"categories":6425},[],{"categories":6427},[5282],{"categories":6429},[5263],{"categories":6431},[5273],{"categories":6433},[],{"categories":6435},[5359],{"categories":6437},[5282],{"categories":6439},[5273],{"categories":6441},[5282],{"categories":6443},[],{"categories":6445},[5273],{"categories":6447},[],{"categories":6449},[5273],{"categories":6451},[],{"categories":6453},[],{"categories":6455},[5279],{"categories":6457},[5314],{"categories":6459},[5299],{"categories":6461},[5273],{"categories":6463},[],{"categories":6465},[5273],{"categories":6467},[5282],{"categories":6469},[5299],{"categories":6471},[5299],{"categories":6473},[5282],{"categories":6475},[5282],{"categories":6477},[5314],{"categories":6479},[5263],{"categories":6481},[],{"categories":6483},[5299],{"categories":6485},[5299],{"categories":6487},[5299],{"categories":6489},[5273],{"categories":6491},[5299],{"categories":6493},[],{"categories":6495},[5279],{"categories":6497},[5299],{"categories":6499},[5273],{"categories":6501},[],{"categories":6503},[5299],{"categories":6505},[],{"categories":6507},[5299],{"categories":6509},[],{"categories":6511},[],{"categories":6513},[],{"categories":6515},[5299],{"categories":6517},[5299],{"categories":6519},[5299],{"categories":6521},[5299],{"categories":6523},[],{"categories":6525},[5299],{"categories":6527},[5299],{"categories":6529},[5299],{"categories":6531},[],{"categories":6533},[5299],{"categories":6535},[],{"categories":6537},[5276],{"categories":6539},[5299],{"categories":6541},[],{"categories":6543},[],{"categories":6545},[],{"categories":6547},[5299],{"categories":6549},[5268],{"categories":6551},[5268],{"categories":6553},[],{"categories":6555},[5273],{"categories":6557},[5299],{"categories":6559},[],{"categories":6561},[5299],{"categories":6563},[5299],{"categories":6565},[5268],{"categories":6567},[],{"categories":6569},[5299],{"categories":6571},[5268],{"categories":6573},[5273],{"categories":6575},[5299],{"categories":6577},[],{"categories":6579},[],{"categories":6581},[],{"categories":6583},[5273],{"categories":6585},[5273],{"categories":6587},[5273],{"categories":6589},[5273],{"categories":6591},[5299],{"categories":6593},[5279],{"categories":6595},[5279],{"categories":6597},[5273],{"categories":6599},[5273],{"categories":6601},[5314],{"categories":6603},[5711],{"categories":6605},[5314],{"categories":6607},[5314],{"categories":6609},[5299],{"categories":6611},[5273],{"categories":6613},[5299],{"categories":6615},[5314],{"categories":6617},[5299],{"categories":6619},[5273],{"categories":6621},[5273],{"categories":6623},[5273],{"categories":6625},[5273],{"categories":6627},[5273],{"categories":6629},[5299],{"categories":6631},[5314],{"categories":6633},[5314],{"categories":6635},[5276],{"categories":6637},[5273],{"categories":6639},[],{"categories":6641},[5273],{"categories":6643},[],{"categories":6645},[5268],{"categories":6647},[5299],{"categories":6649},[],{"categories":6651},[5263],{"categories":6653},[5279],{"categories":6655},[5279],{"categories":6657},[5273],{"categories":6659},[5273],{"categories":6661},[5299],{"categories":6663},[5299],{"categories":6665},[5268],{"categories":6667},[5268],{"categories":6669},[5624],{"categories":6671},[5273],{"categories":6673},[5268],{"categories":6675},[],{"categories":6677},[5299],{"categories":6679},[5273],{"categories":6681},[5273],{"categories":6683},[5273],{"categories":6685},[5273],{"categories":6687},[5299],{"categories":6689},[5299],{"categories":6691},[5299],{"categories":6693},[5299],{"categories":6695},[5273],{"categories":6697},[5273],{"categories":6699},[5273],{"categories":6701},[5273],{"categories":6703},[],{"categories":6705},[5279],{"categories":6707},[5299],{"categories":6709},[5299],{"categories":6711},[5299],{"categories":6713},[],{"categories":6715},[5276],{"categories":6717},[],{"categories":6719},[5314],{"categories":6721},[],{"categories":6723},[5273],{"categories":6725},[5314],{"categories":6727},[5279],{"categories":6729},[5314],{"categories":6731},[],{"categories":6733},[5314],{"categories":6735},[5314],{"categories":6737},[],{"categories":6739},[5279],{"categories":6741},[5273],{"categories":6743},[5273],{"categories":6745},[5314],{"categories":6747},[5299],{"categories":6749},[5299],{"categories":6751},[],{"categories":6753},[5268],{"categories":6755},[],{"categories":6757},[5276],{"categories":6759},[],{"categories":6761},[5279],{"categories":6763},[5268],{"categories":6765},[5279],{"categories":6767},[5279],{"categories":6769},[5279],{"categories":6771},[5279],{"categories":6773},[5279],{"categories":6775},[5279],{"categories":6777},[5279],{"categories":6779},[5279],{"categories":6781},[5279],{"categories":6783},[5279],{"categories":6785},[],{"categories":6787},[5273],{"categories":6789},[5279],{"categories":6791},[5299],{"categories":6793},[5299],{"categories":6795},[5279],{"categories":6797},[5279],{"categories":6799},[5279],{"categories":6801},[5279],{"categories":6803},[5279],{"categories":6805},[5279],{"categories":6807},[5279],{"categories":6809},[5299,5279],{"categories":6811},[5279],{"categories":6813},[5279],{"categories":6815},[5279],{"categories":6817},[5279],{"categories":6819},[],{"categories":6821},[5279],{"categories":6823},[5279],{"categories":6825},[5279],{"categories":6827},[5279],{"categories":6829},[5279],{"categories":6831},[5279],{"categories":6833},[5279],{"categories":6835},[5279],{"categories":6837},[5279],{"categories":6839},[5279,5299],{"categories":6841},[5279],{"categories":6843},[5279],{"categories":6845},[],{"categories":6847},[5268],{"categories":6849},[],{"categories":6851},[5299],{"categories":6853},[],{"categories":6855},[5273],{"categories":6857},[5624],{"categories":6859},[5711],{"categories":6861},[5273],{"categories":6863},[5273],{"categories":6865},[],{"categories":6867},[5273],{"categories":6869},[],{"categories":6871},[5273],{"categories":6873},[],{"categories":6875},[],{"categories":6877},[5299],{"categories":6879},[5299],{"categories":6881},[5299],{"categories":6883},[5268],{"categories":6885},[5268],{"categories":6887},[5268],{"categories":6889},[5268],{"categories":6891},[],{"categories":6893},[5268],{"categories":6895},[],{"categories":6897},[5268],{"categories":6899},[5299],{"categories":6901},[5268],{"categories":6903},[5268],{"categories":6905},[5268],{"categories":6907},[5268],{"categories":6909},[5299],{"categories":6911},[5268],{"categories":6913},[5273],{"categories":6915},[],{"categories":6917},[5273],{"categories":6919},[5268],{"categories":6921},[5299],{"categories":6923},[5268],{"categories":6925},[5268],{"categories":6927},[5268],{"categories":6929},[5299],{"categories":6931},[5299],{"categories":6933},[5299],{"categories":6935},[],{"categories":6937},[],{"categories":6939},[5299],{"categories":6941},[5268],{"categories":6943},[],{"categories":6945},[5299],{"categories":6947},[5273],{"categories":6949},[5299],{"categories":6951},[5273],{"categories":6953},[5273],{"categories":6955},[5299],{"categories":6957},[],{"categories":6959},[],{"categories":6961},[5273],{"categories":6963},[5273],{"categories":6965},[5273],{"categories":6967},[5273],{"categories":6969},[5273],{"categories":6971},[5273],{"categories":6973},[5273],{"categories":6975},[5273],{"categories":6977},[],{"categories":6979},[5273],{"categories":6981},[5273],{"categories":6983},[5273],{"categories":6985},[5299],{"categories":6987},[5299],{"categories":6989},[5299],{"categories":6991},[5268],{"categories":6993},[5299],{"categories":6995},[5299],{"categories":6997},[5299],{"categories":6999},[5273],{"categories":7001},[5276],{"categories":7003},[5276],{"categories":7005},[5276],{"categories":7007},[5273],{"categories":7009},[],{"categories":7011},[5299],{"categories":7013},[],{"categories":7015},[],{"categories":7017},[5299],{"categories":7019},[],{"categories":7021},[5273],{"categories":7023},[5279],{"categories":7025},[5314],{"categories":7027},[5359],{"categories":7029},[5299],{"categories":7031},[5273],{"categories":7033},[5279],{"categories":7035},[],{"categories":7037},[5273],{"categories":7039},[5276,5263],{"categories":7041},[5273],{"categories":7043},[5273],{"categories":7045},[5624],{"categories":7047},[5282],{"categories":7049},[5276],{"categories":7051},[5314],{"categories":7053},[5299],{"categories":7055},[],{"categories":7057},[5299],{"categories":7059},[],{"categories":7061},[5299],{"categories":7063},[5299],{"categories":7065},[5273],{"categories":7067},[],{"categories":7069},[5299],{"categories":7071},[5273],{"categories":7073},[5299],{"categories":7075},[5314],{"categories":7077},[5273],{"categories":7079},[5299],{"categories":7081},[5299,5314],{"categories":7083},[5314],{"categories":7085},[],{"categories":7087},[5299],{"categories":7089},[5299],{"categories":7091},[5299],{"categories":7093},[],{"categories":7095},[],{"categories":7097},[5273],{"categories":7099},[5276],{"categories":7101},[5268],{"categories":7103},[5273],{"categories":7105},[5299],{"categories":7107},[5268],{"categories":7109},[],{"categories":7111},[5314],{"categories":7113},[5268],{"categories":7115},[],{"categories":7117},[5359],{"categories":7119},[5276],{"categories":7121},[5263],{"categories":7123},[5268],{"categories":7125},[5299],{"categories":7127},[5273],{"categories":7129},[5299],{"categories":7131},[5273],{"categories":7133},[5273],{"categories":7135},[5268],{"categories":7137},[5314],{"categories":7139},[5279],{"categories":7141},[5263],{"categories":7143},[5299],{"categories":7145},[5299],{"categories":7147},[],{"categories":7149},[],{"categories":7151},[5299],{"categories":7153},[],{"categories":7155},[5299],{"categories":7157},[5268],{"categories":7159},[],{"categories":7161},[5273],{"categories":7163},[5314],{"categories":7165},[5268],{"categories":7167},[5314],{"categories":7169},[5273],{"categories":7171},[5299],{"categories":7173},[],{"categories":7175},[5273],{"categories":7177},[5273],{"categories":7179},[5279],{"categories":7181},[5273],{"categories":7183},[5279],{"categories":7185},[5273],{"categories":7187},[5273],{"categories":7189},[5279],{"categories":7191},[],{"categories":7193},[],{"categories":7195},[5279],{"categories":7197},[5279],{"categories":7199},[5279],{"categories":7201},[5282],{"categories":7203},[5314],{"categories":7205},[5314],{"categories":7207},[5273],{"categories":7209},[5268],{"categories":7211},[5314],{"categories":7213},[5314],{"categories":7215},[5276],{"categories":7217},[5279],{"categories":7219},[5273],{"categories":7221},[5273],{"categories":7223},[5299],{"categories":7225},[5314],{"categories":7227},[5299],{"categories":7229},[],{"categories":7231},[5624],{"categories":7233},[5711],{"categories":7235},[],{"categories":7237},[],{"categories":7239},[5273],{"categories":7241},[5268],{"categories":7243},[5276],{"categories":7245},[5276],{"categories":7247},[5359],{"categories":7249},[5279],{"categories":7251},[5359],{"categories":7253},[5359],{"categories":7255},[5273],{"categories":7257},[],{"categories":7259},[],{"categories":7261},[5359],{"categories":7263},[5282],{"categories":7265},[5299],{"categories":7267},[5282],{"categories":7269},[5359],{"categories":7271},[5282],{"categories":7273},[5359],{"categories":7275},[5263],{"categories":7277},[5282],{"categories":7279},[5314],{"categories":7281},[5299],{"categories":7283},[],{"categories":7285},[5359],{"categories":7287},[5624],{"categories":7289},[],{"categories":7291},[5299],{"categories":7293},[5299],{"categories":7295},[],{"categories":7297},[],{"categories":7299},[5299],{"categories":7301},[5299],{"categories":7303},[5268],{"categories":7305},[5299],{"categories":7307},[],{"categories":7309},[5268],{"categories":7311},[],{"categories":7313},[],{"categories":7315},[5268],{"categories":7317},[5268],{"categories":7319},[5299],{"categories":7321},[5299],{"categories":7323},[5299],{"categories":7325},[5299],{"categories":7327},[5299],{"categories":7329},[5299],{"categories":7331},[5276],{"categories":7333},[],{"categories":7335},[5299],{"categories":7337},[],{"categories":7339},[],{"categories":7341},[5273],{"categories":7343},[5314],{"categories":7345},[],{"categories":7347},[5624],{"categories":7349},[5299,5624],{"categories":7351},[5299],{"categories":7353},[],{"categories":7355},[5279],{"categories":7357},[5279],{"categories":7359},[5279],{"categories":7361},[5279],{"categories":7363},[5279],{"categories":7365},[],{"categories":7367},[],{"categories":7369},[],{"categories":7371},[5282],{"categories":7373},[5273],{"categories":7375},[5263],{"categories":7377},[5282],{"categories":7379},[5314],{"categories":7381},[5279],{"categories":7383},[],{"categories":7385},[5276],{"categories":7387},[5711],{"categories":7389},[5359],{"categories":7391},[5359],{"categories":7393},[5359],{"categories":7395},[5314],{"categories":7397},[5711],{"categories":7399},[5314],{"categories":7401},[],{"categories":7403},[5263],{"categories":7405},[5282],{"categories":7407},[5299],{"categories":7409},[5279],{"categories":7411},[5276],{"categories":7413},[5282],{"categories":7415},[5276],{"categories":7417},[5299],{"categories":7419},[5279],{"categories":7421},[5282],{"categories":7423},[5624],{"categories":7425},[5299],{"categories":7427},[5268],{"categories":7429},[5282],{"categories":7431},[],{"categories":7433},[5299],{"categories":7435},[5282],{"categories":7437},[5282],{"categories":7439},[5273],{"categories":7441},[],{"categories":7443},[5276],{"categories":7445},[5276],{"categories":7447},[5276],{"categories":7449},[5273],{"categories":7451},[5299],{"categories":7453},[],{"categories":7455},[5263],{"categories":7457},[5314],{"categories":7459},[5314],{"categories":7461},[5359],{"categories":7463},[5263],{"categories":7465},[5268],{"categories":7467},[5359],{"categories":7469},[],{"categories":7471},[5268],{"categories":7473},[5268],{"categories":7475},[5268],{"categories":7477},[5299],{"categories":7479},[5263],{"categories":7481},[5299],{"categories":7483},[],{"categories":7485},[],{"categories":7487},[],{"categories":7489},[5282],{"categories":7491},[5273],{"categories":7493},[],{"categories":7495},[5314],{"categories":7497},[5279],{"categories":7499},[],{"categories":7501},[5276],{"categories":7503},[],{"categories":7505},[5279],{"categories":7507},[5299],{"categories":7509},[5314],{"categories":7511},[5263],{"categories":7513},[],{"categories":7515},[5279],{"categories":7517},[5279],{"categories":7519},[5299],{"categories":7521},[],{"categories":7523},[],{"categories":7525},[5282],{"categories":7527},[5299],{"categories":7529},[],{"categories":7531},[5273],{"categories":7533},[5299],{"categories":7535},[],{"categories":7537},[5282],{"categories":7539},[5273],{"categories":7541},[5299],{"categories":7543},[5359],{"categories":7545},[5299],{"categories":7547},[],{"categories":7549},[5359],{"categories":7551},[5299],{"categories":7553},[5282],{"categories":7555},[5299],{"categories":7557},[5359],{"categories":7559},[5273],{"categories":7561},[5299],{"categories":7563},[5299],{"categories":7565},[5299,5273],{"categories":7567},[5273],{"categories":7569},[5273],{"categories":7571},[5273],{"categories":7573},[5279],{"categories":7575},[5314],{"categories":7577},[5299],{"categories":7579},[5314],{"categories":7581},[5279],{"categories":7583},[5299],{"categories":7585},[],{"categories":7587},[],{"categories":7589},[5299],{"categories":7591},[5299],{"categories":7593},[5299],{"categories":7595},[5273],{"categories":7597},[5299],{"categories":7599},[],{"categories":7601},[5299],{"categories":7603},[5299],{"categories":7605},[5273],{"categories":7607},[5273],{"categories":7609},[5299],{"categories":7611},[5299],{"categories":7613},[],{"categories":7615},[5299],{"categories":7617},[],{"categories":7619},[5299],{"categories":7621},[5299],{"categories":7623},[5299],{"categories":7625},[5299],{"categories":7627},[5299],{"categories":7629},[5299],{"categories":7631},[5299],{"categories":7633},[],{"categories":7635},[5299],{"categories":7637},[5268],{"categories":7639},[5268],{"categories":7641},[],{"categories":7643},[],{"categories":7645},[5299],{"categories":7647},[],{"categories":7649},[5299],{"categories":7651},[5299,5624],{"categories":7653},[],{"categories":7655},[5268],{"categories":7657},[],{"categories":7659},[5299],{"categories":7661},[],{"categories":7663},[],{"categories":7665},[],{"categories":7667},[5299],{"categories":7669},[],{"categories":7671},[5299],{"categories":7673},[],{"categories":7675},[5299],{"categories":7677},[5299],{"categories":7679},[],{"categories":7681},[],{"categories":7683},[5299,5624],{"categories":7685},[5624,5299],{"categories":7687},[5268],{"categories":7689},[],{"categories":7691},[5299],{"categories":7693},[],{"categories":7695},[5299],{"categories":7697},[5299],{"categories":7699},[],{"categories":7701},[5268],{"categories":7703},[5299,5263],{"categories":7705},[5268],{"categories":7707},[5282],{"categories":7709},[],{"categories":7711},[5273],{"categories":7713},[5299],{"categories":7715},[5276],{"categories":7717},[5299],{"categories":7719},[5314],{"categories":7721},[5314],{"categories":7723},[5624],{"categories":7725},[5268],{"categories":7727},[5299],{"categories":7729},[5624],{"categories":7731},[5282],{"categories":7733},[5299],{"categories":7735},[5314],{"categories":7737},[],{"categories":7739},[5299],{"categories":7741},[],{"categories":7743},[],{"categories":7745},[5299],{"categories":7747},[],{"categories":7749},[5299],{"categories":7751},[5282],{"categories":7753},[5263],{"categories":7755},[5314],{"categories":7757},[5276],{"categories":7759},[5273],{"categories":7761},[5314],{"categories":7763},[],{"categories":7765},[5276],{"categories":7767},[],{"categories":7769},[],{"categories":7771},[5299],{"categories":7773},[5268],{"categories":7775},[5276],{"categories":7777},[],{"categories":7779},[5299],{"categories":7781},[5268],{"categories":7783},[5268],{"categories":7785},[5276],{"categories":7787},[5268],{"categories":7789},[5299],{"categories":7791},[5268],{"categories":7793},[5299],{"categories":7795},[],{"categories":7797},[5299],{"categories":7799},[5299],{"categories":7801},[5299],{"categories":7803},[5268],{"categories":7805},[],{"categories":7807},[],{"categories":7809},[5279],{"categories":7811},[5268],{"categories":7813},[],{"categories":7815},[5299],{"categories":7817},[5299],{"categories":7819},[5299],{"categories":7821},[5299],{"categories":7823},[5299],{"categories":7825},[5299],{"categories":7827},[5299],{"categories":7829},[5299],{"categories":7831},[5299],{"categories":7833},[5276],{"categories":7835},[5299,5279],{"categories":7837},[5268],{"categories":7839},[5268],{"categories":7841},[5299],{"categories":7843},[5282],{"categories":7845},[5359],{"categories":7847},[5299],{"categories":7849},[5299],{"categories":7851},[],{"categories":7853},[],{"categories":7855},[5299],{"categories":7857},[5299],{"categories":7859},[],{"categories":7861},[5279],{"categories":7863},[5279],{"categories":7865},[5314],{"categories":7867},[5299],{"categories":7869},[5314],{"categories":7871},[5299],{"categories":7873},[5299],{"categories":7875},[],{"categories":7877},[5299],{"categories":7879},[],{"categories":7881},[],{"categories":7883},[5299],{"categories":7885},[],{"categories":7887},[],{"categories":7889},[5268],{"categories":7891},[],{"categories":7893},[5299],{"categories":7895},[5299],{"categories":7897},[5299],{"categories":7899},[],{"categories":7901},[5299],{"categories":7903},[5268],{"categories":7905},[5711],{"categories":7907},[5273],{"categories":7909},[5299],{"categories":7911},[],{"categories":7913},[5273],{"categories":7915},[5299],{"categories":7917},[],{"categories":7919},[5299],{"categories":7921},[],{"categories":7923},[5273],{"categories":7925},[],{"categories":7927},[],{"categories":7929},[5273],{"categories":7931},[5273],{"categories":7933},[5273],{"categories":7935},[5299],{"categories":7937},[],{"categories":7939},[5273],{"categories":7941},[5273],{"categories":7943},[],{"categories":7945},[],{"categories":7947},[5273],{"categories":7949},[5299],{"categories":7951},[5268],{"categories":7953},[5711],{"categories":7955},[5276],{"categories":7957},[],{"categories":7959},[5279],{"categories":7961},[5299],{"categories":7963},[5299],{"categories":7965},[5263],{"categories":7967},[5268],{"categories":7969},[5268],{"categories":7971},[5268],{"categories":7973},[5268],{"categories":7975},[],{"categories":7977},[5273],{"categories":7979},[5273],{"categories":7981},[5273],{"categories":7983},[5273],{"categories":7985},[5314],{"categories":7987},[5299],{"categories":7989},[5263],{"categories":7991},[],{"categories":7993},[5314],{"categories":7995},[5273],{"categories":7997},[5279],{"categories":7999},[5279],{"categories":8001},[5279],{"categories":8003},[5279],{"categories":8005},[5279],{"categories":8007},[5279],{"categories":8009},[5299,5263],{"categories":8011},[5273],{"categories":8013},[5263],{"categories":8015},[5268],{"categories":8017},[5268],{"categories":8019},[5314],{"categories":8021},[],{"categories":8023},[],{"categories":8025},[5276],{"categories":8027},[],{"categories":8029},[5299],{"categories":8031},[5276],{"categories":8033},[5299],{"categories":8035},[5282],{"categories":8037},[5273],{"categories":8039},[5263],{"categories":8041},[5273],{"categories":8043},[5282],{"categories":8045},[5314],{"categories":8047},[5273],{"categories":8049},[],{"categories":8051},[5314],{"categories":8053},[],{"categories":8055},[],{"categories":8057},[5273],{"categories":8059},[5273],{"categories":8061},[5273],{"categories":8063},[5299],{"categories":8065},[5299],{"categories":8067},[5299],{"categories":8069},[5299],{"categories":8071},[5299],{"categories":8073},[],{"categories":8075},[5624],{"categories":8077},[5299],{"categories":8079},[],{"categories":8081},[],{"categories":8083},[],{"categories":8085},[5314],{"categories":8087},[],{"categories":8089},[5299],{"categories":8091},[],{"categories":8093},[5268],{"categories":8095},[5299],{"categories":8097},[5268],{"categories":8099},[5299],{"categories":8101},[5273],{"categories":8103},[],{"categories":8105},[5299],{"categories":8107},[5299],{"categories":8109},[],{"categories":8111},[5359],{"categories":8113},[5359],{"categories":8115},[5282],{"categories":8117},[5279],{"categories":8119},[],{"categories":8121},[5299],{"categories":8123},[5273],{"categories":8125},[],{"categories":8127},[],{"categories":8129},[5299],{"categories":8131},[5282],{"categories":8133},[5273],{"categories":8135},[5263],{"categories":8137},[5314,5282],{"categories":8139},[5282],{"categories":8141},[5299],{"categories":8143},[5273],{"categories":8145},[],{"categories":8147},[],{"categories":8149},[],{"categories":8151},[],{"categories":8153},[],{"categories":8155},[],{"categories":8157},[5299],{"categories":8159},[],{"categories":8161},[],{"categories":8163},[5299],{"categories":8165},[],{"categories":8167},[],{"categories":8169},[],{"categories":8171},[5299],{"categories":8173},[5268],{"categories":8175},[],{"categories":8177},[],{"categories":8179},[],{"categories":8181},[5299],{"categories":8183},[],{"categories":8185},[5299],{"categories":8187},[5299],{"categories":8189},[],{"categories":8191},[5299],{"categories":8193},[5282],{"categories":8195},[],{"categories":8197},[5314],{"categories":8199},[5314],{"categories":8201},[],{"categories":8203},[5276],{"categories":8205},[],{"categories":8207},[],{"categories":8209},[],{"categories":8211},[5279],{"categories":8213},[5268],{"categories":8215},[5273],{"categories":8217},[5299],{"categories":8219},[5263],{"categories":8221},[5299],{"categories":8223},[],{"categories":8225},[],{"categories":8227},[5263],{"categories":8229},[5276],{"categories":8231},[5273],{"categories":8233},[],{"categories":8235},[5624],{"categories":8237},[],{"categories":8239},[5276],{"categories":8241},[5299],{"categories":8243},[5299],{"categories":8245},[5276],{"categories":8247},[5299],{"categories":8249},[5279],{"categories":8251},[5273],{"categories":8253},[5299],{"categories":8255},[5273],{"categories":8257},[5299],{"categories":8259},[5273],{"categories":8261},[5314],{"categories":8263},[5314],{"categories":8265},[5279],{"categories":8267},[],{"categories":8269},[5299],{"categories":8271},[5299],{"categories":8273},[5276],{"categories":8275},[5711],{"categories":8277},[5314],{"categories":8279},[5268],{"categories":8281},[5299],{"categories":8283},[5268],{"categories":8285},[5299],{"categories":8287},[5299],{"categories":8289},[],{"categories":8291},[5299],{"categories":8293},[],{"categories":8295},[5299],{"categories":8297},[5276],{"categories":8299},[5299],{"categories":8301},[5299],{"categories":8303},[5299],{"categories":8305},[],{"categories":8307},[5299],{"categories":8309},[5299],{"categories":8311},[5711],{"categories":8313},[],{"categories":8315},[5268],{"categories":8317},[5624],{"categories":8319},[5282],{"categories":8321},[],{"categories":8323},[5359],{"categories":8325},[],{"categories":8327},[],{"categories":8329},[5268],{"categories":8331},[5299],{"categories":8333},[],{"categories":8335},[5299],{"categories":8337},[5299],{"categories":8339},[5273],{"categories":8341},[5299],{"categories":8343},[5268],{"categories":8345},[5268],{"categories":8347},[5279],{"categories":8349},[5279],{"categories":8351},[5279],{"categories":8353},[5299],{"categories":8355},[5359],{"categories":8357},[5268],{"categories":8359},[5314],{"categories":8361},[],{"categories":8363},[5279],{"categories":8365},[5279],{"categories":8367},[5624],{"categories":8369},[5279],{"categories":8371},[5279],{"categories":8373},[5273],{"categories":8375},[5268],{"categories":8377},[5624],{"categories":8379},[5299],{"categories":8381},[5299],{"categories":8383},[5299],{"categories":8385},[5299],{"categories":8387},[],{"categories":8389},[5273],{"categories":8391},[5299],{"categories":8393},[5279],{"categories":8395},[],{"categories":8397},[],{"categories":8399},[5268],{"categories":8401},[],{"categories":8403},[5273],{"categories":8405},[5273],{"categories":8407},[5273],{"categories":8409},[5273],{"categories":8411},[5273],{"categories":8413},[5273],{"categories":8415},[5273],{"categories":8417},[5273],{"categories":8419},[],{"categories":8421},[],{"categories":8423},[5299],{"categories":8425},[],{"categories":8427},[5273],{"categories":8429},[5314],{"categories":8431},[5314],{"categories":8433},[5359],{"categories":8435},[5263],{"categories":8437},[],{"categories":8439},[],{"categories":8441},[],{"categories":8443},[5279],{"categories":8445},[5299],{"categories":8447},[],{"categories":8449},[5263],{"categories":8451},[5263],{"categories":8453},[5279],{"categories":8455},[5314],{"categories":8457},[5359],{"categories":8459},[5279],{"categories":8461},[5279],{"categories":8463},[],{"categories":8465},[5273],{"categories":8467},[5263],{"categories":8469},[5263],{"categories":8471},[5299],{"categories":8473},[5273],{"categories":8475},[5282],{"categories":8477},[5279],{"categories":8479},[],{"categories":8481},[5276],{"categories":8483},[5359],{"categories":8485},[5268],{"categories":8487},[5268],{"categories":8489},[5268],{"categories":8491},[5624],{"categories":8493},[],{"categories":8495},[5273],{"categories":8497},[],{"categories":8499},[5273],{"categories":8501},[5273],{"categories":8503},[5299],{"categories":8505},[5299],{"categories":8507},[5282],{"categories":8509},[5273],{"categories":8511},[5282],{"categories":8513},[],{"categories":8515},[5273],{"categories":8517},[5279],{"categories":8519},[5279],{"categories":8521},[5279],{"categories":8523},[5299],{"categories":8525},[5273],{"categories":8527},[5299],{"categories":8529},[5263],{"categories":8531},[5268],{"categories":8533},[5279],{"categories":8535},[5268],{"categories":8537},[5299],{"categories":8539},[],{"categories":8541},[5268],{"categories":8543},[5273],{"categories":8545},[5268],{"categories":8547},[5268],{"categories":8549},[5268],{"categories":8551},[5268],{"categories":8553},[],{"categories":8555},[],{"categories":8557},[5268],{"categories":8559},[5268],{"categories":8561},[],{"categories":8563},[5268],{"categories":8565},[5268],{"categories":8567},[5299],{"categories":8569},[5299],{"categories":8571},[5268],{"categories":8573},[5268],{"categories":8575},[5299],{"categories":8577},[],{"categories":8579},[5299],{"categories":8581},[5273],{"categories":8583},[5299],{"categories":8585},[5299],{"categories":8587},[],{"categories":8589},[5299],{"categories":8591},[5299],{"categories":8593},[5299],{"categories":8595},[5268],{"categories":8597},[],{"categories":8599},[],{"categories":8601},[],{"categories":8603},[],{"categories":8605},[5299],{"categories":8607},[5299],{"categories":8609},[],{"categories":8611},[5276],{"categories":8613},[5268],{"categories":8615},[],{"categories":8617},[],{"categories":8619},[],{"categories":8621},[],{"categories":8623},[],{"categories":8625},[5299],{"categories":8627},[],{"categories":8629},[],{"categories":8631},[5299],{"categories":8633},[],{"categories":8635},[5273],{"categories":8637},[5273],{"categories":8639},[5273],{"categories":8641},[5263],{"categories":8643},[],{"categories":8645},[5276],{"categories":8647},[5282],{"categories":8649},[5282],{"categories":8651},[5624],{"categories":8653},[5268],{"categories":8655},[],{"categories":8657},[5299],{"categories":8659},[5299],{"categories":8661},[5263],{"categories":8663},[],{"categories":8665},[5263],{"categories":8667},[],{"categories":8669},[],{"categories":8671},[],{"categories":8673},[5282],{"categories":8675},[5273],{"categories":8677},[5273],{"categories":8679},[5273],{"categories":8681},[5273],{"categories":8683},[5273],{"categories":8685},[],{"categories":8687},[5268],{"categories":8689},[5299],{"categories":8691},[5299],{"categories":8693},[5299],{"categories":8695},[],{"categories":8697},[5263],{"categories":8699},[],{"categories":8701},[5279],{"categories":8703},[5359],{"categories":8705},[5279],{"categories":8707},[],{"categories":8709},[],{"categories":8711},[5299],{"categories":8713},[5273],{"categories":8715},[],{"categories":8717},[5299],{"categories":8719},[5299],{"categories":8721},[5299],{"categories":8723},[5273],{"categories":8725},[5273],{"categories":8727},[5299],{"categories":8729},[5359],{"categories":8731},[5273],{"categories":8733},[],{"categories":8735},[5299],{"categories":8737},[],{"categories":8739},[5711],{"categories":8741},[5282],{"categories":8743},[5359],{"categories":8745},[5282],{"categories":8747},[5624],{"categories":8749},[5299],{"categories":8751},[5282],{"categories":8753},[5268],{"categories":8755},[5624],{"categories":8757},[5282],{"categories":8759},[5279],{"categories":8761},[5279],{"categories":8763},[],{"categories":8765},[5282],{"categories":8767},[],{"categories":8769},[5314],{"categories":8771},[5282],{"categories":8773},[],{"categories":8775},[5359],{"categories":8777},[5359],{"categories":8779},[5711],{"categories":8781},[],{"categories":8783},[5299],{"categories":8785},[5282],{"categories":8787},[5624],{"categories":8789},[5273],{"categories":8791},[5273],{"categories":8793},[5359],{"categories":8795},[5299],{"categories":8797},[5314],{"categories":8799},[5299],{"categories":8801},[],{"categories":8803},[],{"categories":8805},[],{"categories":8807},[5276],{"categories":8809},[5299],{"categories":8811},[5279],{"categories":8813},[5282],{"categories":8815},[5282],{"categories":8817},[5299],{"categories":8819},[5276],{"categories":8821},[5314],{"categories":8823},[5299],{"categories":8825},[5282],{"categories":8827},[5299],{"categories":8829},[5282],{"categories":8831},[5314],{"categories":8833},[5314],{"categories":8835},[5273],{"categories":8837},[5314],{"categories":8839},[5282],{"categories":8841},[5263],{"categories":8843},[5282],{"categories":8845},[5282],{"categories":8847},[5282],{"categories":8849},[5282],{"categories":8851},[],{"categories":8853},[5268],{"categories":8855},[],{"categories":8857},[5359],{"categories":8859},[5299],{"categories":8861},[5299],{"categories":8863},[],{"categories":8865},[],{"categories":8867},[],{"categories":8869},[5299],{"categories":8871},[5268],{"categories":8873},[5299],{"categories":8875},[5299],{"categories":8877},[],{"categories":8879},[5299],{"categories":8881},[5279],{"categories":8883},[5299],{"categories":8885},[5299],{"categories":8887},[5299],{"categories":8889},[],{"categories":8891},[],{"categories":8893},[],{"categories":8895},[5624],{"categories":8897},[5624],{"categories":8899},[5263],{"categories":8901},[5273],{"categories":8903},[5263,5276],{"categories":8905},[5299],{"categories":8907},[5268],{"categories":8909},[],{"categories":8911},[5279],{"categories":8913},[5359],{"categories":8915},[5299],{"categories":8917},[5282],{"categories":8919},[5299],{"categories":8921},[],{"categories":8923},[5359],{"categories":8925},[5624],{"categories":8927},[5273],{"categories":8929},[5263],{"categories":8931},[5624],{"categories":8933},[5273],{"categories":8935},[5314],{"categories":8937},[5273],{"categories":8939},[5314],{"categories":8941},[5299],{"categories":8943},[5314],{"categories":8945},[5314],{"categories":8947},[5282],{"categories":8949},[5359],{"categories":8951},[5299],{"categories":8953},[5276],{"categories":8955},[],{"categories":8957},[5299],{"categories":8959},[5279],{"categories":8961},[5359],{"categories":8963},[5263],{"categories":8965},[5299],{"categories":8967},[5359],{"categories":8969},[5314],{"categories":8971},[5299],{"categories":8973},[5299],{"categories":8975},[5359],{"categories":8977},[5299],{"categories":8979},[5314],{"categories":8981},[5299],{"categories":8983},[],{"categories":8985},[5299],{"categories":8987},[5299],{"categories":8989},[5299],{"categories":8991},[5299],{"categories":8993},[],{"categories":8995},[5273],{"categories":8997},[5624],{"categories":8999},[],{"categories":9001},[],{"categories":9003},[5299],{"categories":9005},[5263],{"categories":9007},[5276],{"categories":9009},[5263],{"categories":9011},[5263],{"categories":9013},[5273],{"categories":9015},[],{"categories":9017},[5299],{"categories":9019},[5268],{"categories":9021},[5299],{"categories":9023},[5299],{"categories":9025},[],{"categories":9027},[5273],{"categories":9029},[5268],{"categories":9031},[5299,5624],{"categories":9033},[5273,5624],{"categories":9035},[5624],{"categories":9037},[5299],{"categories":9039},[5273],{"categories":9041},[5273],{"categories":9043},[5282],{"categories":9045},[5282],{"categories":9047},[5282],{"categories":9049},[5299],{"categories":9051},[5279],{"categories":9053},[5273],{"categories":9055},[],{"categories":9057},[5624],{"categories":9059},[],{"categories":9061},[5624],{"categories":9063},[5624],{"categories":9065},[5263],{"categories":9067},[5273],{"categories":9069},[],{"categories":9071},[5624],{"categories":9073},[5299],{"categories":9075},[5268],{"categories":9077},[5299],{"categories":9079},[5279],{"categories":9081},[5282],{"categories":9083},[5282],{"categories":9085},[5282],{"categories":9087},[5624],{"categories":9089},[],{"categories":9091},[],{"categories":9093},[],{"categories":9095},[5299],{"categories":9097},[5282],{"categories":9099},[5299],{"categories":9101},[5282],{"categories":9103},[5624],{"categories":9105},[5624],{"categories":9107},[5299],{"categories":9109},[5273],{"categories":9111},[],{"categories":9113},[5299],{"categories":9115},[5299],{"categories":9117},[5299],{"categories":9119},[],{"categories":9121},[],{"categories":9123},[5624],{"categories":9125},[5624],{"categories":9127},[5299,5624],{"categories":9129},[5273],{"categories":9131},[5273],{"categories":9133},[5273],{"categories":9135},[5273],{"categories":9137},[5273],{"categories":9139},[5273],{"categories":9141},[],{"categories":9143},[5282],{"categories":9145},[5299],{"categories":9147},[5282],{"categories":9149},[5276],{"categories":9151},[5299],{"categories":9153},[5711],{"categories":9155},[5711],{"categories":9157},[5273],{"categories":9159},[5282],{"categories":9161},[],{"categories":9163},[5273],{"categories":9165},[5299],{"categories":9167},[],{"categories":9169},[5279],{"categories":9171},[],{"categories":9173},[5299],{"categories":9175},[5273],{"categories":9177},[5268],{"categories":9179},[5299],{"categories":9181},[],{"categories":9183},[],{"categories":9185},[5279],{"categories":9187},[5279],{"categories":9189},[5314],{"categories":9191},[5279],{"categories":9193},[5273],{"categories":9195},[],{"categories":9197},[5273],{"categories":9199},[5268],{"categories":9201},[5299],{"categories":9203},[5299],{"categories":9205},[],{"categories":9207},[5299],{"categories":9209},[5314],{"categories":9211},[5299],{"categories":9213},[],{"categories":9215},[5359],{"categories":9217},[5282],{"categories":9219},[5282],{"categories":9221},[5263],{"categories":9223},[5263],{"categories":9225},[5263],{"categories":9227},[5273],{"categories":9229},[5263],{"categories":9231},[5273],{"categories":9233},[5624],{"categories":9235},[5711],{"categories":9237},[5268],{"categories":9239},[5268],{"categories":9241},[5268],{"categories":9243},[5624],{"categories":9245},[5268,5263],{"categories":9247},[5359],{"categories":9249},[5273],{"categories":9251},[],{"categories":9253},[5299],{"categories":9255},[],{"categories":9257},[5282],{"categories":9259},[5359],{"categories":9261},[5279],{"categories":9263},[5282],{"categories":9265},[5314],{"categories":9267},[],{"categories":9269},[5273],{"categories":9271},[],{"categories":9273},[5711],{"categories":9275},[],{"categories":9277},[5279],{"categories":9279},[5279],{"categories":9281},[5359],{"categories":9283},[],{"categories":9285},[5299],{"categories":9287},[5359],{"categories":9289},[],{"categories":9291},[5299],{"categories":9293},[5299],{"categories":9295},[],{"categories":9297},[5314],{"categories":9299},[5299],{"categories":9301},[],{"categories":9303},[5299],{"categories":9305},[],{"categories":9307},[],{"categories":9309},[5273],{"categories":9311},[5273],{"categories":9313},[],{"categories":9315},[5282],{"categories":9317},[5282],{"categories":9319},[5282],{"categories":9321},[5299,5273],{"categories":9323},[5273],{"categories":9325},[5273],{"categories":9327},[5273],{"categories":9329},[5359],{"categories":9331},[5359],{"categories":9333},[],{"categories":9335},[5268],{"categories":9337},[5299],{"categories":9339},[5359],{"categories":9341},[5359],{"categories":9343},[5268],{"categories":9345},[5263],{"categories":9347},[5273],{"categories":9349},[5282],{"categories":9351},[5299],{"categories":9353},[5299],{"categories":9355},[5273],{"categories":9357},[5282],{"categories":9359},[5273],{"categories":9361},[5299],{"categories":9363},[5276],{"categories":9365},[],{"categories":9367},[5299],{"categories":9369},[],{"categories":9371},[5299],{"categories":9373},[5299],{"categories":9375},[5282],{"categories":9377},[],{"categories":9379},[5359],{"categories":9381},[5299],{"categories":9383},[5273],{"categories":9385},[5273],{"categories":9387},[5282],{"categories":9389},[5314],{"categories":9391},[5314],{"categories":9393},[5268],{"categories":9395},[5299],{"categories":9397},[5273],{"categories":9399},[],{"categories":9401},[5273],{"categories":9403},[5299],{"categories":9405},[5268],{"categories":9407},[5299],{"categories":9409},[5299],{"categories":9411},[5299],{"categories":9413},[5273],{"categories":9415},[5359],{"categories":9417},[5299],{"categories":9419},[5279],{"categories":9421},[5299],{"categories":9423},[5299],{"categories":9425},[5299],{"categories":9427},[5299],{"categories":9429},[],{"categories":9431},[5299],{"categories":9433},[5359],{"categories":9435},[5279],{"categories":9437},[5299],{"categories":9439},[5279],{"categories":9441},[],{"categories":9443},[],{"categories":9445},[],{"categories":9447},[5299],{"categories":9449},[],{"categories":9451},[],{"categories":9453},[],{"categories":9455},[],{"categories":9457},[5273],{"categories":9459},[5314],{"categories":9461},[5273],{"categories":9463},[5273],{"categories":9465},[5282],{"categories":9467},[5263],{"categories":9469},[5299],{"categories":9471},[5299],{"categories":9473},[5299],{"categories":9475},[5263],{"categories":9477},[5314],{"categories":9479},[],{"categories":9481},[5359],{"categories":9483},[5276],{"categories":9485},[5299],{"categories":9487},[5279],{"categories":9489},[5314],{"categories":9491},[5314],{"categories":9493},[5711],{"categories":9495},[5273],{"categories":9497},[5299],{"categories":9499},[5299],{"categories":9501},[5314],{"categories":9503},[5299],{"categories":9505},[],{"categories":9507},[],{"categories":9509},[5624],{"categories":9511},[5279],{"categories":9513},[5314],{"categories":9515},[5299],{"categories":9517},[5268],{"categories":9519},[5314],{"categories":9521},[5263],{"categories":9523},[5273],{"categories":9525},[5273],{"categories":9527},[5268],{"categories":9529},[5299],{"categories":9531},[],{"categories":9533},[],{"categories":9535},[],{"categories":9537},[5299],{"categories":9539},[],{"categories":9541},[5268],{"categories":9543},[],{"categories":9545},[5299],{"categories":9547},[],{"categories":9549},[5268],{"categories":9551},[5273],{"categories":9553},[5299],{"categories":9555},[5624],{"categories":9557},[5299],{"categories":9559},[5314],{"categories":9561},[5299],{"categories":9563},[5314],{"categories":9565},[5314],{"categories":9567},[],{"categories":9569},[],{"categories":9571},[5314],{"categories":9573},[5314],{"categories":9575},[5314],{"categories":9577},[],{"categories":9579},[5314],{"categories":9581},[5273],{"categories":9583},[5273],{"categories":9585},[],{"categories":9587},[5299],{"categories":9589},[5276],{"categories":9591},[5359],{"categories":9593},[5299],{"categories":9595},[],{"categories":9597},[5314],{"categories":9599},[5299],{"categories":9601},[5711],{"categories":9603},[5314],{"categories":9605},[5314],{"categories":9607},[5276],{"categories":9609},[5282],{"categories":9611},[5282],{"categories":9613},[],{"categories":9615},[5282],{"categories":9617},[5299],{"categories":9619},[],{"categories":9621},[],{"categories":9623},[5273],{"categories":9625},[],{"categories":9627},[5273],{"categories":9629},[5273],{"categories":9631},[5268],{"categories":9633},[5299],{"categories":9635},[5268],{"categories":9637},[5314],{"categories":9639},[5268],{"categories":9641},[5282],{"categories":9643},[5282],{"categories":9645},[5282],{"categories":9647},[5268],{"categories":9649},[5299],{"categories":9651},[5273],{"categories":9653},[5624],{"categories":9655},[5263],{"categories":9657},[5624],{"categories":9659},[5624],{"categories":9661},[5282],{"categories":9663},[5624],{"categories":9665},[5624]]