Archon V3: YAML Harnesses for AI Coding Agents

Archon V3 replaces 8 manual AI coding steps (classify, investigate, plan, implement, review, test, commit, PR) with one YAML command, using Git worktrees for 4+ parallel isolated runs, DAGs for parallelism, and hooks for self-correction—enabling Stripe-scale output (1,300 PRs/week) without babysitting.

Harness Engineering Unlocks Reliable AI Coding at Scale

Manual AI coding fails due to five issues: inconsistent outputs from same prompts, context bloat in long sessions causing hallucinations, no parallelism (one agent/repo/task), fear of delegation without oversight, and non-composable skills/commands rebuilt per task. Archon V3 introduces "harness engineering"—the layer beyond prompt and context engineering—turning these into deterministic YAML workflows that mix precise steps, creative AI nodes, and loops until tests pass.

Stripe ships 1,300 PRs/week with zero human code; OpenAI hit 3.5 PRs/engineer/day on a million-line project using the same models via harnesses. Encode workflows as YAML committed to repos for team sharing/forking. Information flows via artifact files (not chat history), keeping sessions sharp under 200k tokens. Run from CLI/Web/Slack/GitHub/Discord/Telegram with Claude Code or Codex SDKs, mixing providers per-node to avoid lock-in.

Three Primitives: Commands, DAG Workflows, Git Worktree Isolation

Commands are single-task Markdown files (e.g., classify issue) with frontmatter for variables like {args}, {artifacts_dir}, {workflow_id}. Keep to one job for reusability across workflows.

Workflows define directed acyclic graphs (DAGs) in YAML: nodes declare dependencies/conditions (e.g., code review + security review run parallel post-classify; branch bug-fix vs. feature on classified.output.type == "bug"). Arkon schedules parallelism automatically.

Isolation via Git worktrees: each run gets a fresh ~/.arkon/workspaces/<id> directory/branch/sandbox. Run 4+ workflows parallel (bugfix, feature, review, refactor) without conflicts; main repo untouched. List with arkon isolation list; auto-cleanup >7 days old or arkon isolation cleanup.

User-level ~/.arkon holds DB/worktrees/config; repo-level .arkon (git-committed) has custom commands/workflows. Install in 60s: curl script (Mac/Linux), PowerShell (Windows), brew install, or Docker. First run: arkon workflow run archon-assist "your question".

Hooks and Built-in Workflows for Self-Correction and Production Speed

PreToolUse hooks (before tool call): inject context, deny writes (e.g., review nodes can't edit files), rewrite inputs—all in YAML, not prompts.

PostToolUse hooks (after): enable loops like "reread your write, verify type-checks, rewrite if needed" for self-correcting quality without prompt tweaks. Reliability from feedback wiring, not better prompts.

Built-ins (forkable YAML):

archon assist: Q&A/exploration.
archon fix-github-issue: classify/investigate/implement/review/PR.
archon idea-to-pr: paragraph → reviewed PR.
archon smart-pr-review: scales to complexity.
archon comprehensive-pr-review: parallel multi-reviewer.
archon architect: simplify hotspots.
archon conflict-resolution: full-repo merge fixes.

Future: visual builder, more workflows/SDKs/hooks. 91% solo AI builders quit in 3 months without community; join for daily hangs/workshops.

Video description

Archon V3 is live — the first open source harness builder for AI coding agents. Encode any Claude Code or Codex workflow as YAML, run it from CLI, Web, Slack, GitHub, or Discord, and replace eight manual steps with one command. Prompt engineering, context engineering, and now harness engineering — the next layer for shipping real code with AI. ---- 🚀 Want to learn agentic coding with live daily events and workshops? Check out Dynamous AI: https://dynamous.ai/?code=646a60 Get 10% off here 👉 https://shorturl.smartcode.diy/dynamous_ai_10_percent_discount ---- Chapters 0:00 Archon V3 Preview: YAML Workflows, Worktrees, Six Adapters 0:09 Harness Engineering: The Next Layer After Prompt and Context 0:59 What Archon Actually Is: First Open Source Harness Builder for AI Coding 1:48 Stripe 1,300 PRs/Week and OpenAI 3.5 PRs/Engineer/Day — Why the Harness Matters 2:55 Three Primitives: Commands, Workflows, and Isolation 3:28 Archon Architecture: User-Level vs Repo-Level, How Artifacts Replace Chat History 4:37 Git Worktrees: Run Four AI Workflows in Parallel Without Conflicts 5:32 Install Archon in 60 Seconds (Mac, Linux, Windows, Homebrew, Docker) 6:42 DAG Workflows: Parallel Code Review + Security Review in One Run 7:44 PreToolUse and PostToolUse Hooks: Self-Correcting Quality Loops 8:46 Built-In Production Workflows: archon fix-github-issue, idea-to-PR, Smart PR Review 9:43 Writing Your First Custom Command — Turn Skills and Slash Commands Into Workflows 10:26 Six Adapters: CLI, Web, Slack, Discord, Telegram, GitHub — Plus Claude + Codex SDKs 11:19 Where Archon Goes Next: Visual Workflow Builder, More SDKs, Deeper Hooks Resources ⭐ Archon on GitHub: https://github.com/coleam00/Archon 📖 The Archon Book: https://archon.diy/book 🎓 Dynamous AI Community: https://dynamous.ai/?code=646a60 💰 10% OFF Dynamous: https://shorturl.smartcode.diy/dynamous_ai_10_percent_discount Key Concepts Covered Harness Engineering — The evolution from prompt engineering and context engineering. A harness is the system around the coding agent that turns manually shepherding eight steps every day into one command. Deterministic steps where you need precision, AI steps where you need creativity, loops that iterate until the tests actually pass. YAML Workflows as Code — Archon workflows are YAML files committed to your repo. Read them, fork them, bend them to your team's exact process. The workflow is the contract between you and the agent. DAG Execution and Parallelism — Describe your workflow as a directed acyclic graph. Archon figures out which nodes can run in parallel, which depend on which, and what conditions gate runtime branches. Code review and security review run at the same time. Bug-fix and feature paths branch on classification output. Git Worktrees for Isolation — Every workflow run gets its own worktree, its own branch, its own sandbox. Four parallel workflows, zero conflicts, your main checkout never notices. Stop babysitting, start dispatching. PreToolUse and PostToolUse Hooks — Inject context, deny calls, rewrite inputs, or build self-correcting quality loops. Your code review node writes nothing. Your implementation node reviews its own writes before moving on. Reliability comes from wiring feedback, not from writing better prompts. Ecosystem and Adapters Archon ships with production-ready built-in workflows you can run the moment you install it: - archon assist — questions and exploration - archon fix-github-issue — classify, investigate, implement, review, open PR - archon idea-to-PR — full pipeline from one-paragraph description to reviewed pull request - archon smart-pr-review — reviewers scale to complexity - archon comprehensive-pr-review — parallel multi-reviewer analysis - archon architect — finds complexity hotspots and simplifies them - archon conflict-resolution — merge messes handled end-to-end Run any of them from six different surfaces: CLI, Web UI, Slack, Discord, Telegram, or GitHub. Mix Claude Code SDK and Codex SDK per-node inside the same DAG. Multi-provider, not multi-vendor lock-in. About This Channel DIY Smart Code — deep dives on AI coding tools, agentic engineering, Claude Code, Codex, open source developer tools, and real workflows from the community. If you're building real software with AI agents and you want the honest technical breakdown (not hype), subscribe. --- So — harness engineering, the next evolution, or are you sticking with raw Claude Code and manual steps? Drop your take below. If you want Archon updates as they ship, hit subscribe — the visual workflow builder is landing next and I'll break it down the same way. #ArchonV3 #HarnessEngineering #AICoding #ClaudeCode #Codex #AgenticCoding #OpenSource #YAMLWorkflows #DAGWorkflows #GitWorktrees #AIAgents #DeveloperTools #ClaudeCodeSDK #CodexSDK #CodingAgents #AIWorkflows #AgentEngineering #ColeMedin #Dynamous #AIAutomation #PromptEngineering #ContextEngineering #SelfCorrectingAgents #DevTools #Archon