On November 11, 2026, Anthropic published research titled "Worse Together" on its Alignment Science blog: multi-user, multi-agent collaboration has a systematic weakness — when every person sends their own agent to compete for the same shared resource, the group outcome is worse than assigning a single coordinator agent to arrange things for everyone. The tested models came from Anthropic, OpenAI, and Alibaba, and three of the environments have been released as a new benchmark, MAMUBench.
Where the coordinator wins: all four tests, by 10 to 45 percentage points
The researchers built four simulated environments: researchers sharing one API-token budget, patients sharing a clinic calendar, several people assembling a group food order or travel booking through their own assistants, and multiple merge requests queuing for a release. The models tested included Claude Opus 5, Claude Sonnet 5, GPT-5.6, and Qwen3.8-max. The coordinator won in all four environments: for Claude Opus 5, measured as a share of the best achievable outcome, the API-budget scenario came out 64 to 30, the clinic calendar 84 to 71, the personal assistant 88 to 44, and the merge queue 92 to 81 — a lead of 10 to 45 percentage points.
Where agent teams lose: they act too late, and decide for others without asking
The problem is not that individual agents lack ability, but collaboration itself. As teams grew from 4 to 16 agents, the share that ever took action kept falling: Claude Sonnet 5 dropped from 66% to 10%, and GPT-5.6-sol from 82% to 25%. Worse still was deciding for others without asking: GPT and Qwen agents cancelled or shrank other users' jobs without consent an average of 3.9 to 7.1 times per episode, against 0.59 for Claude Sonnet 5 and zero for Opus 5. Group orders had a race condition too: the special request reached the agent placing the order in 94% to 99% of episodes, but was actually read by checkout time in only 47% to 67%, so the chance of the request being honored fell from 56% to 68% down to 19% to 36%. More than half of Claude team episodes also contained claims with no basis in the evidence, versus 17% to 20% for a single agent. An earlier test on this site reached a similar conclusion: Multi-Agent Teams Put to the Test: Costs Up to 5.1x Higher, Quality Barely Moved.
Fixes depend on the setting, and platforms have a part to play
The fixes that work depend on the setting. In the API-budget environment, appointing a team lead with decision authority, ranking jobs by value, and releasing budget promptly brought the team back to coordinator level; in the group-order setting the same approach barely helped, and the most effective fix was a platform guardrail — blocking checkout until the agent had read all unread messages, which recovered 73% (Opus 5) and 62% (Sonnet 5) of failed episodes. These settings have already left the lab, and the industry is building a shared language for them, for example the A2A Protocol. MAMUBench points in a clear direction: more agents does not mean better coordination, and without a coordinator role or platform guardrails, extra agents may trip each other up first.
