Claude Fable 5.2 Formalizes Fermat With 1,000 Agents

This is less a story about a model proving mathematics alone than a case study in agent infrastructure: shared state, verification, and dynamic coordination made the scale-up work. It is impressive but bounded—Anthropic calls it a toy experiment and seeded the team with 47 previously proved lemmas.

• –Shared Postgres tables and Git branches let 1,000 agents claim tasks, exchange notes, and submit independently checked Lean proofs.
• –Lean’s compiler provided an objective gate for progress, sharply reducing the risk of plausible-looking but invalid work.
• –Sandboxing, no-network execution, command classifiers, and offline monitoring show that large agent teams require security architecture as much as model capability.
• –For developers, the key lesson is to invest in state management, task allocation, verification, and recovery—not simply more agents.