Safely Scaling a 1,000-Agent Team to Reprove Fermat’s Last Theorem

Anthropic reports a toy experiment in which a team of 1,000 agents running an early version of Claude Fable 5.2 reproduced a formal proof of Fermat’s Last Theorem without the domain-specific platform and researcher steering used in its earlier work. The September formalization relied on Prove2Me, a shared graph that linked statements to their sub-statements, and took dozens of agents 11 days and roughly 6 billion output tokens, producing 13 million lines of Lean code. In the new run, the agents used only a shell, a file editor, Lean‘s toolchain, Git, and a shared Postgres database. The team took 28.7 hours, used one-third as many output tokens, and wrote a proof one-seventh the length of the original.

The harness gave each agent its own sandboxed virtual machine with no network access, connected only through a gateway that relayed Claude API calls and Git and SQL requests. A prompted LLM classifier checked every shell command and file edit before it ran, and an offline monitor reviewed the team’s actions afterward. The agents were given 60 statements to prove, a one-sentence sketch of the classical route, definitions, and 47 proved lemmas from the earlier formalization. The only organizational structure was a hierarchy: one lead and 999 helpers. The lead was told to split the problem and assign tasks; helpers were told to expect work from the lead.

Within the first 10 minutes, the lead created five database tables: Announce, Assign, Notes, Subtasks, and Landed. It assigned each statement to a crew of 16-17 helpers, designated a captain per crew, and posted instructions. Helpers initially judged 47 of the 60 statements infeasible within the budget, but the lead reassigned nearly all helpers to the 10 statements they thought finishable, and all 10 were proved within roughly 70 minutes. The lead then moved helpers to the harder statements, and by hour three told free helpers not to wait for assignments.

Around hour 18, only Langlands-Tunnell, a classical result Wiles’s proof relies on, remained along with dependent statements. Its captain had estimated at least 0.7-1 million lines of Lean code and less than a 1% chance of finishing in time. The lead refocused about 750 helpers on it. Because many initial tasks were too large for one helper, the team organized work as a tree that grew to 2,418 tasks up to 10 levels deep. Langlands-Tunnell was proved at hour 28.4, and the final theorem at hour 28.7.

Safety combined sandboxing, real-time monitoring, and offline monitoring. Sandboxes had no network access, database calls ran under an unprivileged role with a five-second time limit and 256 MiB write cap, and kernels were built without features giving root a direct path in. The real-time classifier blocked tool calls that tried to escape the sandbox, probe boundaries, or tamper with monitoring; any block stopped the whole team until a person reviewed it. Across more than a million tool calls in this run, none met the bar to be blocked, though replaying them against a chain-of-thought monitor and today’s rubric brought one root-access check close. The run used only the action-based monitor; a parallel CoT monitor was added shortly afterward.

The authors conclude that domain-specific harnesses like Prove2Me may still improve efficiency, that human direction can help when agents misjudge difficulty or give up, and that carefully designed safety boundaries were what made the experiment possible. They expect multi-agent teams to take on harder work and believe such teams, scaled safely, will help solve challenging problems.

Safely scaling multi-agent teams with Claude Fable 5.2

View Original