Anthropic's thousand-agent harness: 11 days of work in 28.7 hours
blaizedsouza · x · 2026-11-14
Anthropic's "Thousand-Agent Harness" writeup explains how Fable 5.2 compressed an 11-day task (formalizing Fermat's Last Theorem) into 28.7 hours at roughly 3× fewer tokens: the gain came from better splitting, not a smarter agent.
- Architecture: one lead agent that only splits and assigns work (never writes the fix), with up to 999 helpers running in parallel in their own sandboxes and git branches
- Coordination: five shared database tables (orders/assignments/progress/taskqueue/landed) replace agent-to-agent messaging, with 750 agents reading/writing concurrently; the helper prompt is identical for all 999
- Verification: a merge + grader gate counts only verified, landed work; the resulting proof is 7× shorter, with no special platform or researcher hints
- Shrinks to Claude Code: a lead hands out tasks, helpers take them in their own git worktrees, and one tasks-and-status file keeps everyone in sync
The scale rule: a bigger agent doesn't finish faster, a better-split one does — give every agent one task, one sandbox, and one place to report.
Related event: 1,000 Claude Agents Team Up to Formally Prove Fermat's Last Theorem(3 posts)→
More from coding & agent
- Weaviate Podcast: Lightweight reranking can cut agent token costs despite slower search — CShorten30 · 2026-11-14
- Slashvibe launches $189 physical device for managing multiple AI coding agents — Scobleizer · 2026-11-14
- Gemini agent conferences critique quantum code, 290x cheaper than best published version — alex_verem · 2026-11-14
- OpenAI agents can now buy things via ChatGPT Wallet with biometric approval per purchase — rohanpaul_ai · 2026-11-14
- Grady Booch asks if Microsoft's egregious outages correlate with its embrace of AI-driven development — Grady_Booch · 2026-11-14
- Coding agents: effort levels will vanish, replaced by budgets like "$10 and 30 minutes" — johnlindquist · 2026-11-14