Shared notes let small agent teams beat larger search pools

Researchers at Microsoft Research and UC Berkeley report that language-model agents can solve some test-time search tasks more efficiently by exchanging results during a run. In their paper, a few communicating agents matched or exceeded much larger pools of independent attempts. The findings suggest that shared state and verifiable intermediate scores can reduce the brute-force sampling needed for open-ended problems.

Each experiment gives identical agents the same objective and per-agent compute budget. The agents work through a shared workspace without predefined roles or a central orchestrator, exchanging partial results, failed approaches, code, and other artifacts while the search continues. The paper compares team@k, a group of k communicating agents, with best@k, the highest-scoring result from k independent runs.

Five agents match 33 solo attempts

ARC-AGI-3 tests agents on unfamiliar grid-world games whose rules must be inferred through interaction. Performance improved as the communicating group grew: team@3 matched best@13, while team@5 matched best@33. Those teams delivered the same results as 4.3 and 6.6 times as many independent agents, respectively.

Individual games showed larger differences. Across 64 independent attempts at LP85, none succeeded, while team@5 reached a 65% solve rate. On FT09, team@3 increased the solve rate from about 26% to 90%.