Anthropic: 1,000 Claude Fable 5.2 agents reproduce Fermat proof formalization in 28.7 hours
The team used one-third the output tokens of the September run and wrote a proof one-seventh the length, with no Prove2Me platform and no researcher steering, Anthropic says.
What happened
Anthropic ran 1,000 AI agents as one team to rebuild a computer-checkable proof of Fermat's Last Theorem. Each agent worked in its own locked-down sandbox with no internet. The agents shared work through a database and Git, a tool for tracking code changes.
Why it matters to you
It shows AI agent teams may handle very large projects with little human direction, if safety limits are built in around them.
The takeaway
Anthropic says large agent teams can do long, hard tasks inside carefully designed safety boundaries.
What it means for you
Anthropic reports that a shared SQL database plus per-agent virtual machines let 1,000 agents coordinate, after a single shared machine hit limits at about 100 agents.
The source describes a research experiment, not a product, so it gives no direct guidance for business use yet.