Agent Labs: Playbook for High-Growth AI Startups
Agent Labs build agents over models, using product-first strategies, outcome pricing up to $2000/month, and human-in-loop control to achieve better economics and PMF than capital-intensive Model Labs.
Agent Labs Embed Business Plans in Agent-Focused Products
Agent Labs like Cursor ($29B valuation), Perplexity ($20B), Cognition ($10B), Sierra ($10B), Lovable ($2B), and Gamma ($2B) succeed by researching and selling agents, not models. Unlike Neolabs (e.g., Thinking Machines, SSI) that chase overlooked model research, Agent Labs follow a repeatable playbook: product-first development, outcome-based pricing, user-centric autonomy, and cost-aware evals.
Start with proven products—Cursor forked VSCode first, then iterated on models after 2 years of user insights—avoiding Magic.dev's $100M bet on unproven long-context models. Charge for outcomes ($2000/month possible) to gain pricing power and margins by replacing human labor, escaping Model Labs' 9-900x annual distillation grind and token-based pricing limits. Prioritize speed, auditable human-in-the-loop control, and multiturn interactivity over raw autonomy hours; rewrite harnesses frequently for gains. Focus evals on intelligence/success vs. cost Pareto frontiers for high-volume usage, not just max capabilities like IMO/IOI benchmarks.
Conway's Law reveals priorities: Agent Labs allocate resources to full-stack delivery engineers (FDEs) and GTMEs as top talent, paying applied AI engineers 50-70% less than research staff in Model Labs. They open-source agents (e.g., OpenAI's sales/support/research assistants; Vercel's 5 agents for support to data analysis) to commoditize complements, abstract model selection into task models, and sweat B2B needs. High acquihire retention signals product focus.
Model Labs Pivot to Platforms, Unlocking Agent Lab Season
Model Labs dedicate <30% of compute to inference (OpenAI's case per Epoch AI), with most resources on unpublished research. Products like Operator, NotebookLM Audio, and Deep Research get abandoned. Now, OpenAI signals AI Cloud and third-party apps (quoting Bill Gates Line), prioritizing hyperscaler scale down-stack (chips/datacenters) over up-stack superapps. Anthropic unifies Claude Developer efforts amid $350B fundraise and $50B datacenter; Vercel/GitHub/Cloudflare (acquiring Replicate) follow suit.
This blesses Agent Labs: increased frontier model diversity (US/China open labs) means users pay specialists to "capabilitymaxx" model-harness combos full-time. Agents bundle model+prompt+memories+tools+planning—orchestration+auth, eroding model-only moats. Pretraining nears limits after 7-13 years (AlexNet 2012, GPT-1 2018); RL era favors domain focus. Agent Labs like Cursor/Cognition start from open weights, using continued training/post-training (Cursor's log-scale gains close open-to-frontier gap) to match/exceed best humans.
Bear Case and Lasting Fork Potential
Retain "Labs" for R&D in agent engineering/research—fast experimentation beyond tax writeoffs. Bear case: embedded Agent Labs (Claude Code $1B ARR, Codex, Google Labs) dominate, forking model trees. One-size-fits-all AGI (GPT-4o omnimodality) falters; GPT-5 router issues, gpt-5-codex persistence, and "Moving Beyond One-Size-Fits-All" signal task-specialized models, rewarding Agent Labs' domain depth over generalists.