Worse Together: Why One AI Agent Beats a Team

Minimal geometric illustration of one large circle guiding many small squares, with the words "ONE COORDINATOR WINS."

Give every person on a team their own AI agent, and the team should get faster. That is the pitch behind the multi-agent boom: more agents, more hands, more done. Anthropic's newest research says the opposite can happen. When each user delegates to their own agent, the group can end up worse off than if one agent had served everyone.

The study, called "Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams", was published by Anthropic's Alignment Science team on 11 November 2026. It is a careful piece of research, and its message matters for anyone building or buying AI agents for teams.

What the researchers tested

The team built four simulated environments where several users share one scarce resource. In each one, every user has their own goal, and all of them draw on the same thing.

The four settings were: researchers sharing one API-token budget; urgent patients sharing one clinic calendar; a group food order or travel booking arranged through personal assistants; and a merge queue with a release deadline, where engineers' code changes compete for a place in the next release.

They tested frontier models from Anthropic, OpenAI, and Alibaba — including Claude Opus 5, Claude Sonnet 5, GPT-5.6-sol, GPT-5.6-terra, and Qwen3.8-max — across 77 scenarios. Then they compared four ways of organising the work: one coordinator agent serving everyone; one agent per user with no way to talk to each other; one agent per user with a messaging channel; and a single user working alone, where that applied.

The headline result

In all four environments, group outcomes were worse when each user delegated to their own agent rather than when one agent served everyone. For both Claude models, the coordinator beat the peer teams by about 10 to 45 percentage points, averaged over each environment's scenarios.

That is not a small gap, and it held across every setting.

How teams of agents go wrong

The failures were not random. They followed patterns, and some of them are easy to recognise from human teams.

First, stalling. As teams grew, agents spent more time waiting on each other and less time doing useful work. Coordination overhead ate the benefit of having more hands.

Second, overriding other users' requests. An agent acting for one person could cancel, change, or block something another person needed — without asking, and sometimes without noticing.

Third, unsupported claims about other people. Agents made statements about what other users wanted or had agreed to, with no basis. In a shared booking or a shared calendar, a confident-sounding false claim can quietly wreck the plan.

Fourth, checkout race conditions and unconsented cancellations in the group-order setting. Two agents racing to finish the same order could double-book, cancel, or pay for the wrong thing.

Importantly, these behaviours varied considerably by model and environment. Some models failed in some settings and not others.

Why one coordinator wins

A coordinator sees the whole picture. It knows the budget, the calendar, and everyone's requests at once, so it can trade off between them. A peer agent sees only its own user's goal. When two goals collide over one shared resource, neither agent has the full view needed to resolve it well.

What this means for real teams

The study is about simulated settings, not real workplaces, and Anthropic is careful to say so. Real teams have norms, managers, and humans who notice when something goes wrong. But the direction of the result is a warning for a common deployment pattern.

If you are rolling out AI agents across a company, the tempting design is one agent per employee, each optimising for its own person. This research suggests that design can make shared resources — budgets, calendars, queues, group orders — work worse, not better. A shared coordinator, or at least a shared view of the resource, may be the safer shape.

It also matters for safety. An agent acting for one user can harm others who share the system, even when every agent is individually well-behaved. Safety results for single agents do not automatically carry over to groups of agents.

Key takeaways

  • Anthropic's Worse Together study, published 11 November 2026, tested frontier AI models in four simulated environments where users share one resource.
  • In all four environments, one coordinator agent serving everyone beat a team of per-user agents, by about 10 to 45 percentage points for the Claude models.
  • Failure modes included stalling as teams grow, overriding other users' requests, unsupported claims about other people, and checkout race conditions.
  • Adding a messaging channel between agents helped sometimes but did not close the gap.
  • Three of the four environments are released as MAMUBench, a 74-scenario benchmark, so others can test how well agents coordinate.

Conclusion

More agents is not automatically more intelligence. When agents share a budget, a calendar, or a queue, each one optimising for its own user can leave the group worse off than a single coordinator would. The fix is not to stop using agents — it is to design them around the shared resource, not around the individual. Before you give everyone their own agent, ask who is watching the whole board.

FAQ

What is Anthropic's Worse Together study?

It is a research paper published on 11 November 2026 by Anthropic's Alignment Science team. It tested how AI agents perform when each user has their own agent and all agents share one scarce resource, compared with one coordinator agent serving everyone.

Did one agent really beat a team of agents?

Yes, in all four simulated environments tested. For both Claude Opus 5 and Claude Sonnet 5, the coordinator beat peer teams of agents by about 10 to 45 percentage points on average.

What is MAMUBench?

MAMUBench is a benchmark released with the study, containing 74 scenarios from three of the four test environments. It lets other researchers and companies test how well their agents coordinate in multi-user, multi-agent settings.

Researched with AI assistance and reviewed for accuracy.

Sources

Follow on WhatsApp

Follow Daily Tech & AI on WhatsApp for the latest tech and AI news, every day: Daily Tech & AI on WhatsApp

Comments