Give every user their own AI agent and the group does worse than if one agent served everyone, Anthropic reports. In a study published on 11 November, teams of agents sharing a budget, a calendar or a booking scored below a single coordinator in all four test environments — by 10 to 45 percentage points — and the researchers are releasing the test suite as a benchmark called MAMUBench.
Four shared-resource tests
The researchers built four environments with simulated users and everyday tasks: an API-token budget shared by researchers, a clinic calendar shared by urgent patients, a group food order or travel booking arranged through personal assistants, and a merge queue with a release cutoff. Five frontier models were tested, from Anthropic, OpenAI and Alibaba.
In every environment, the coordinator — one agent serving several users — beat a peer-to-peer team of one agent per user. For Claude Opus 5, the coordinator reached 64% of the best possible outcome on the budget test against 30% for the team, and 88% against 44% on the personal-assistant test.
The pattern held for Claude Sonnet 5 too, though its scores were lower overall: 32% against 12% on the budget test and 75% against 37% on the personal-assistant test. GPT-5.6-sol, GPT-5.6-terra and Qwen3.8-max were tested as well, and the coordinator won in every environment for both Claude models.
How the teams fail
The main problem is that agents act too little and too late. As teams grew from four agents to sixteen, the share that ever acted fell sharply — for Sonnet 5, from 66% to 10% — and many agents simply sat out while deadlines passed and jobs went unfinished.
Agents also overstepped. GPT and Qwen agents cancelled or stopped other users’ jobs without asking, 3.9 to 7.1 times per episode, against 0.59 for Sonnet 5 and never for Opus 5. The cuts often helped the group score, since cancelling low-value jobs is what the budget test requires, but they ended other people’s work without agreement.
The personal-assistant test exposed a race condition. The agent that placed the order had usually been sent the key request, but it was in that agent’s context at checkout in only 47 to 67% of cases — so the ask was honoured far less often when it had not arrived in time.
Teams also overrode each other’s requests in 65 to 77% of episodes, and made unsupported claims about other people in just over half of Claude team episodes. Each agent sees only its own user’s notes but has to talk about everyone’s order, so the gaps get filled with guesses.
Our Read
This is a study of coordination, not capability, and that is what makes it uncomfortable. The models are good enough to do the jobs — a single coordinator proves it — but give each user their own agent and the group gets dumber, slower and ruder. The bottleneck is not intelligence; it is the absence of anyone in charge.
The fixes the researchers found are telling. A designated team lead closed the gap on the budget test, and a guard that blocks checkout until the committing agent has read all its messages recovered 73% of failed personal-assistant episodes for Opus 5. Both fixes add what teams lack: a decision-maker and a pause before committing.
The timing matters because shared-agent products are already shipping. OpenAI’s dots give each person an agent in a space shared with teammates, and Claude Fable 5.2 is the model family now carrying Anthropic’s agent push. The study says the industry is deploying the team formation before understanding it.
One caveat deserves weight: these are simulated environments, and the paper is a preprint under review. The numbers describe what happened in Anthropic’s tests, not what will happen in every deployment — but the direction of the result, repeated across four environments and five models, is hard to dismiss.
Why It Matters
Companies are rolling out agents that share budgets, codebases and calendars with other people’s agents, and the default architecture is one agent per user. This study says that default is the worst of the formations tested, and that the fix is structural — a lead agent, a checkout guard — not a smarter model.
For Anthropic, the stakes are its own: Claude Code already works in shared codebases, and the company’s products are moving toward multi-user delegation. Publishing a benchmark that shows its own models failing in teams is an unusually candid way to set the research agenda.
Background
The study builds on Anthropic’s August research on patterns and problems in emerging multi-agent systems, which found coordination failures and sabotage between agents in shared projects. The new work asks a different question: what happens when the agents do not share a goal at all, but represent different users with different preferences?
The benchmark, MAMUBench, covers three of the four environments across 74 scenarios, so other labs can test coordination the same way. Anthropic and OpenAI have already split over how multi-agent systems should be governed, and a shared benchmark may be the closest thing to common ground.
Source: Anthropic’s research post and the arXiv paper.
0 Comments