Ten Claude Agents, One Overnight Run, and a 122-Year-Old Math Problem
On September 30, an account began circulating across the Chinese internet describing an unusual overnight experiment: ten instances of Claude Sonnet 5.5, each working on the same goal, spent roughly 15 hours exchanging about 1,270 messages, wrote nearly 17,900 lines of Lean proof code, and reported a result for the N=7 case of the Thomson problem — a question about how electrons arrange themselves on a sphere that has been open since 1904.
What the experiment claims
According to the account, no human directed the work. The agents formed their own group, disagreed with each other, chose algorithms, and merged code on their own. The result is expressed in Lean, a proof assistant, which means it can in principle be checked mechanically rather than taken on faith. That detail is a big part of why the story traveled: a machine-checkable proof is a far stronger claim than a fluent write-up.
How to read it
Treat this as a reported result, not settled science. A Lean proof still has to survive independent review — from mathematicians, and from people checking that the formal statement actually matches the original problem. The more interesting signal is the pattern: several general-purpose agents, a long time horizon, and a formal language acting as the referee. It is a vivid data point for the argument that AI is shifting from "answering questions" toward "doing research," and a natural follow-up to this year's other headline math-and-AI stories.
Why it resonated in China
Chinese tech circles have been unusually focused on AI agents that can run for hours and use tools, a shift often summed up with the meme 养龙虾 — using an AI agent to do your work for you. An overnight, ten-agent "research run" is that idea taken to its logical extreme, and it arrived as the country celebrated its own AI milestones. Even readers who treat the claim cautiously tend to agree on the trend line it points at.
Related vocab: 养龙虾.