Give every user their own AI agent and the group does worse, an Anthropic study finds
A study from Anthropic's Fellows program put agents for different users on one shared budget, calendar or cart. One agent serving everybody won in all four tests, though not by as much as the headline suggests.
Most of the companies selling AI agents are building the same thing: every person gets their own. Your agent books your seat, mine books mine, and somewhere in the middle they split one budget, one calendar or one group order. Anthropic’s Fellows program just tested what happens then, and the answer in its simulations was not good.
The paper, titled “Worse Together,” went up on Anthropic’s Alignment Science blog on November 11. In all four environments the researchers built, group outcomes were worse when each user handed the job to a personal agent than when one agent served everyone. Anthropic’s own models took part, along with models from OpenAI and Alibaba.
What did they actually test?
Four setups, all simulated, each with several fake users competing for one shared resource: researchers splitting an API-token budget that couldn’t cover everyone’s jobs, urgent patients sharing a clinic calendar, friends arranging a group food order or trip booking through assistants, and engineers racing a release cutoff in a merge queue.
They compared a single “coordinator” agent serving all the users against a peer-to-peer team, one agent per user with a channel to message the others. Five models were used: Claude Opus 5, Claude Sonnet 5, GPT-5.6-sol, GPT-5.6-terra and Qwen3.8-max. The headline comparison, per Anthropic, is that the coordinator beat the team in all four environments for both Claude models, by about 10 to 45 percentage points.
How big was the gap, really?
Bigger than a rounding error, smaller than a rout. The winner’s score is the part I’d read twice. In the shared-budget test, Anthropic says the teams reached 47% and 38% of the coordinator’s result for Opus 5 and Sonnet 5. Those are relative figures. Against the best possible outcome, the coordinator scored 64% and 32%, and the teams 30% and 12%.
| Shared-budget test | Claude Opus 5 | Claude Sonnet 5 |
|---|---|---|
| Coordinator, share of best possible outcome | 64% | 32% |
| Peer-to-peer team, share of best possible outcome | 30% | 12% |
My read: a coordinator that gets a third of the best result with Sonnet 5 isn’t a success story, it’s the less bad option. Anthropic says the environments were designed so a frontier model acting as a coordinator usually performs well, so the 32% suggests this particular test was hard.
Where do the teams break down?
Three patterns. First, that agents freeze as groups grow. From four users to 16, the share of agents that took any action at all fell from 66% to 10% for Sonnet 5 and from 82% to 25% for GPT-5.6-sol.
Second, agents override each other. In the shared-budget test, 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. Anthropic notes those cuts didn’t lower the group’s score, since cancelling low-value jobs is what the test required, but nobody agreed to them. It also admits that its own scoring only counted the value of finished jobs, so it couldn’t see that.
Third, they invent things about people. In the group-ordering test, just over half of Claude team episodes (58% for Opus 5, 53% for Sonnet 5, judged by Claude Opus 5) contained at least one claim about another person that the agent’s information didn’t support. A single agent did that in 17 to 20% of episodes. Anthropic’s explanation is simple: each agent sees only its own user’s notes but has to talk about everyone’s order, so it fills gaps, for instance by assuming the table wants what its own user wants.
Why did the group order fail?
This one is almost mundane. One user in each scenario makes a targeted request, such as a user with celiac disease asking for two plain gluten-free sides. Anthropic found the request reached the agent placing the order in 94 to 99% of episodes. It was in that agent’s context when it checked out in only 47 to 67%.
The agent that commits the order often does so before it has read every message sent to it. A race condition, not a reasoning failure. Simple as that. The coordinator honored the request about twice as often as the team.
What fixed it?
Different things in different tests. Anthropic says that suggests the failures have different causes. In the shared-budget environment, a separate team lead agent told to prioritize high-value jobs and use the budget fully closed the gap with the coordinator. In the group-ordering test a lead didn’t help. What did was a platform guard that blocks checkout until the committing agent has read all its pending messages. It recovered 73% of failed Opus 5 episodes and 62% of failed Sonnet 5 episodes.
Anthropic isn’t telling everyone to build a coordinator. They say a single coordinator isn’t always feasible or natural for systems with many users, so the risks can’t be fixed just by centralizing. Three of the four environments are being released as an open benchmark called MAMUBench.
What this doesn’t show
Short version: it is a simulation.
All of it is simulated users and scripted scenarios. Anthropic’s post makes no claim about any shipping product. Some of the judging was done by Claude models, including Claude Opus 5 grading the fabrication count, so Anthropic’s models are partly grading Anthropic’s models. The authors say this is research from the Anthropic Fellows Program, and the paper is on arXiv.
My read: the useful takeaway for anyone shipping agents isn’t “use one agent”. It’s that the boring plumbing mattered more than model quality here. A guard that makes an agent read its messages before it pays sounds like cheap platform logic rather than a research problem. If you’re building group features on top of personal agents, that’s the first fix in the study I’d copy. For how a single agent handles the “ask first” question, see our write-up of Claude Duet, and for agents already checking out for shoppers, Shopify’s agent checkout and Amazon’s block.
I’d watch whether OpenAI or Google respond with their own multi-user numbers. Nothing in the sources says either plans to.
Sources
- Anthropic Alignment Science Blog: Worse Together, How Performance Breaks Down in Multi-User Multi-Agent Teams
- arXiv: Worse Together, How Performance Breaks Down in Multi-User Multi-Agent Teams
- GitHub: MAMUBench, the released benchmark environments
Note on sources: every figure above is from Anthropic’s blog post. I read the blog post in full and the paper’s abstract and summary through a search snippet, and I didn’t run the benchmark myself. No independent outlet’s coverage was used as a source.