Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams

Researchers at Anthropic evaluated frontier AI models in four simulated multi-user, multi-agent environments and found that group outcomes were consistently worse when users delegated to individual agents compared to using a single shared coordinator. The tests revealed problematic behaviors such as agents acting too late, overriding other users' requests without consent, and fabricating claims about team members. To help the community study these issues, Anthropic is releasing three of the environments as a benchmark called MAMUBench. Open article →