1,000 Claude Agents Proved Fermat's Last Theorem in 28.7 Hours — With No Boss and No Special Tools

Anthropic gave 1,000 AI agents one of the hardest problems in mathematics — a formal, computer-checked proof of Fermat's Last Theorem — and almost no instructions on how to do it. The team finished in 28.7 hours, using one-third of the output tokens and writing a proof one-seventh the length of Anthropic's first attempt in September.

No researcher steered the run. No special math platform coordinated the agents. What they had was Git, a shared Postgres database, and a model strong enough to invent its own organization.

The setup: a theorem, 60 statements, and a blank org chart

Fermat's Last Theorem states that no positive integers a, b, and c satisfy aⁿ + bⁿ = cⁿ for any n greater than 2. It went unproven for over 350 years until Andrew Wiles published his 129-page proof in 1995. Formalizing it — translating every step into Lean, a language where a computer checks each step — has been a long-standing goal in mathematics.

Anthropic's first formalization, published in September 2026, took a researcher and dozens of agents 11 days and roughly 6 billion output tokens. They wrote 13 million lines of Lean code, coordinated through Prove2Me, a platform built for formalizing mathematics.

The new experiment stripped all of that away. A team of 1,000 agents running an early version of Claude Fable 5.2 got:

• 60 statements to prove, including the final theorem and 59 intermediate milestones
• A one-sentence sketch of the classical proof route, plus the definitions those statements use
• 47 proved lemmas from the earlier formalization
• A shell and file editor in a sandbox with no internet access, Lean's toolchain and Mathlib, their own Git branch with merge requests into main, and a shared Postgres database

The only structure: one lead agent and 999 helpers. The lead was told to split the problem and assign work; helpers were told to expect assignments and ask when unclear. Nothing else.

The agents invented their own coordination system

Within the first 10 minutes, the lead agent built five tables in the shared database:

• **Announce** — the lead's orders to the whole team
• **Assign** — who works on what
• **Notes** — an open board for claiming work and reporting progress
• **Subtasks** — a to-do list anyone could add to, first claimer owns it
• **Landed** — a log of finished work

Then the lead dealt out the work: 999 rows in the assign table, each statement handed to a crew of 16–17 helpers, with one captain per crew. Helpers first judged whether their statement was feasible — and most concluded it wasn't. About 27 minutes in, three statements were already proved, and the lead set aside the 47 statements helpers had judged infeasible within budget, piling nearly all helpers onto the remaining 10. Crews grew from about 17 helpers to between 46 and 139.

The lead's verdict at 45 minutes: "the team's throughput is far higher than the early 'infeasible-in-budget' verdicts assumed." All 10 statements were proved within roughly 70 minutes.

Fro