Artificial intelligence agents operating without central coordinators produced new mathematical discoveries and proofs across five distinct problems. Automated mathematics previously depended on human guidance or rigid, pre-scripted computational pipelines. Here, software agents from diverse model families selected their own goals, wrote code, exchanged papers in a shared digital library, and verified calculations together.
When set loose in the digital workspace called the Station, individual bots initiated independent computational experiments to explore numerical patterns. The programs posted their intermediate findings to a common forum, much like human researchers sharing draft preprints on an academic server. Peer agents read these posts, flagged logical errors, and modified existing algorithms to extend the mathematical arguments. Beyond raw numerical constructions, the agents drafted formal theorems with explanatory text to show why the geometric patterns hold.
The research team tested the multi-agent system across twelve construction tasks from the AlphaEvolve catalogue alongside two extra case studies. The autonomous collective produced novel mathematical results on five separate problems, including exact 604-point kissing configurations in dimension eleven. The agents also discovered new infinite families for Book Ramsey numbers and improved the lower bound for Erdős’s minimum-overlap problem.
Mathematicians can now inspect the released agent dialogues, verified code, and mathematical proofs to build directly upon the findings. The authors released all raw verification code and conversation transcripts to provide a transparent record of each mathematical derivation.