On November 9, 2026, Anthropic described a multi-agent experiment on its developer blog: 1,000 agents running an early version of Claude Fable 5.2 reproduced the formal proof of Fermat's Last Theorem in just 28.7 hours. In September, the same proof took a researcher plus dozens of agents 11 days and around 6 billion output tokens. This time the output tokens were one-third of September's, the written proof one-seventh as long — and there was no researcher directing the work and no dedicated Prove2Me collaboration platform.
No one in charge — the division of labor grew on its own
The experiment gave the agents only 60 statements to prove, a one-sentence sketch of the classical route, the definitions those statements use, and 47 already-proved lemmas; how to organize was entirely up to them. In the first 10 minutes, the lead agent built five tables in a shared database: orders, assignments, notes, a to-do list, and a log of finished work. After 27 minutes, three statements were already proved. The lead then set aside 47 statements that most helpers judged infeasible within the budget and concentrated everyone on the remaining 10, with crews growing from around 17 helpers to between 46 and 139. Each agent ran in its own lightweight virtual machine with no network access, reaching the Claude API only through a gateway and sharing work through the database and a Git repository.
A thousand agents working at once also needs keeping in check. Anthropic put three layers around the team: sandboxing, real-time monitoring, and offline review. Every shell command and file edit from every agent first passed a classifier checking for attempts to escape the sandbox or tamper with monitoring; a single block stops the whole team until a person reviews it. Across more than a million tool calls in the run, none reached the bar for blocking. Anthropic adds the caution that this was only a 'toy experiment', run on a model version that had not yet completed its full pre-release alignment assessment, so the conclusions cannot be carried straight into production.
What is genuinely interesting here is not that 'AI can prove a math theorem', but that a team of agents with no one in charge built its own coordination structure. In September, the Prove2Me platform and a researcher held it together; this time, general-purpose tools and a shared database table were enough. Multi-agent teams have been put to the test before: an earlier experiment found that a multi-agent team cost 5.1 times as much while quality barely moved. But the cost is plain to see too: 1,000 sandboxes, more than a million calls, 28.7 hours of continuous running — in exchange for a demonstration with clear boundaries. Whether multi-agent systems move from demonstration to everyday productivity will come down to cost, stability, and the ability to stop things when they go wrong.
