Thursday, November 12, 2026 · Claude & Anthropic
Claude · Research · happened Nov 11, 2026
Anthropic researchers find teams of per-user AI agents do worse than one shared agent, and release MAMUBench
A study from the Anthropic Fellows Program tested Claude Opus 5, Sonnet 5, two OpenAI GPT-5.6 models and Alibaba's Qwen3.8-max in four simulated multi-user tasks, such as a shared API budget and a group food order. A single coordinating agent beat teams with one agent per user by about 10 to 45 points for the Claude models, and teams often acted too late, overrode users' requests or made unsupported claims about others; the GPT and Qwen agents also cancelled other users' jobs without asking several times per episode. Three of the environments are released as a benchmark called MAMUBench.
Why it matters
It suggests the always-on agents the labs are shipping may need explicit coordination before they share work across people.