● The finding: giving every person their own agent performs worse than one shared coordinator agent serving the whole group.
● The gap: the coordinator led by roughly 10 to 45 percentage points across four environments.
● The worst case: on a shared API-token budget, peer-to-peer teams reached only 47% of the coordinator’s result with Opus 5, and 38% with Sonnet 5.
● Released: three of the four environments as MAMUBench, with code on GitHub and the paper on arXiv.
Anthropic Fellows researchers published this on 11 November. They tested five frontier models — Claude Opus 5 and Sonnet 5, GPT-5.6-sol and GPT-5.6-terra, and Alibaba’s Qwen3.8-max — across four settings where several people each have an agent and those agents share one resource: an API token budget, a clinic calendar, a group food and travel assistant, and a merge queue with a release cutoff.
| Environment |
Opus 5 coordinator / team |
Sonnet 5 coordinator / team |
| Shared API key |
64 / 30 |
32 / 12 |
| Clinic calendar |
84 / 71 |
74 / 65 |
| Personal assistant |
88 / 44 |
75 / 37 |
| Merge queue |
92 / 81 |
93 / 78 |
Two numbers explain why. As the group grew from 4 to 16 agents, the share that ever took any action at all collapsed — from 66% to 10% for Sonnet 5, and from 82% to 25% for GPT-5.6-sol. Most agents simply stopped participating.
And in the personal assistant setting, team episodes contained fabricated claims about other people in 58% of Opus 5 runs and 53% of Sonnet 5 runs, against 17–20% for a single agent doing the same job. Agents filled gaps in what they knew about other users by inventing it.
Both failures have fixes, and they are cheap. Appointing one agent as team lead closed the gap in the API-key environment. A guard that forces an agent to read unread messages before finishing recovered 73% of failed personal-assistant episodes with Opus 5 and 62% with Sonnet 5. If you are building multi-agent systems, those two patterns are the practical takeaway.
Sources