AI Agent Swarm Cracks Five Open Math Problems in a Coordinator-Free Sandbox

Researchers built a shared environment called the Station where AI agents drawn from different model families work toward a common research goal with no central controller and no predefined pipeline. Left to their own devices, the agents pick their own directions, run experiments, collaborate, and accumulate a common body of results that later agents can build on—closer to how a research community operates than to a single scripted solver.

Run against 12 construction problems from the AlphaEvolve catalogue plus two extra case studies, the system produced results new to the mathematical literature on five of them. These include a fresh infinite family of finite-field Kakeya sets, exact 604-point kissing configurations in 11 dimensions, record values for the discretized Kakeya needle and sign uncertainty problems, and a notably better lower bound for Erdős’s minimum-overlap problem. The agents also turned up new infinite families for Book Ramsey numbers.

What distinguishes the work from prior automated-discovery efforts is that the agents didn’t just spit out numbers—they generated theorems and explanations of why their constructions hold, making the output something mathematicians can verify and extend rather than treat as a black box. The team released the full record: raw agent conversations, proofs, and verification code, offering a transparent trail of how each discovery came about.