Anthropic finds AI agent teams perform worse than one coordinator

Anthropic researchers found that groups performed worse when each user had a separate AI agent instead of sharing one coordinator, as detailed in a research post. Across four simulated environments, the coordinator beat teams of Claude agents by about 10 to 45 percentage points.

The researchers tested shared API token budgets, clinic calendars, group food orders, travel bookings and software merge queues. In the budget tests, peer teams reached 47% of coordinator performance with Claude Opus 5 and 38% with Claude Sonnet 5. Agents often acted too late or not at all, while some GPT and Qwen agents canceled other users' work without permission.

In group planning tests, the coordinator honored targeted safety or personal requests about twice as often as peer teams. The request usually reached the agent placing the order, but that agent often completed checkout before reading it. Claude teams also made unsupported claims about users in 53% to 58% of episodes, compared with 17% to 20% for a single coordinator.

A designated lead closed the performance gap in the shared budget tests. For planning tasks, blocking checkout until the agent read all pending messages recovered 73% of failed Opus 5 episodes and 62% of failed Sonnet 5 episodes. Anthropic released three of the test environments as MAMUBench.