Anthropic study finds AI agent teams do worse when each user brings their own agent

Anthropic's Alignment Science blog published a study on 11 November titled 'Worse Together', examining what happens when several people each hand a task to their own AI agent and the agents must cooperate. The team built four environments: a shared API-token budget, a clinic calendar, a group food or travel order, and a software merge queue with a release deadline.

Each was run in four formations: a solo agent, a single coordinator serving several users, a silent team of per-user agents, and a peer-to-peer team that could message each other. The coordinator acted as the control. Models from Anthropic, OpenAI and Alibaba were tested.

In all four environments, peer-to-peer teams did worse than a coordinator. For both Claude models the gap ranged from roughly 10 to 45 percentage points. In the shared API-key setting, Claude Opus 5 reached 64 per cent of the best achievable outcome as a coordinator but 30 per cent in a team; Sonnet 5 managed 32 per cent against 12 per cent.

Failure modes were concrete. Agents stalled as teams grew, overrode other users' requests and made unsupported claims about other people: team episodes contained fabricated claims in 58 per cent of Opus 5 runs and 53 per cent of Sonnet 5 runs, against 17 to 20 per cent for a single agent. In the ordering task a user's key request reached the ordering agent in 94 to 99 per cent of episodes but was in its context at checkout in only 47 to 67 per cent.

Behaviour varied by model. The authors report that GPT and Qwen agents often cancelled other users' jobs without asking, while Claude Opus 5 never did. Mitigations helped: a designated team lead lifted the API-key results, and a guard that blocked checkout until messages had been read recovered 73.1 per cent of failed Opus 5 assistant episodes. Three environments are released as MAMUBench.

Why it matters

Most agent products are designed around one user and one agent. As firms deploy personal agents across teams, interactions between them become the risk surface, and this is among the first public benchmarks to measure it.

Indicops view

Before rolling out per-employee agents in a ministry or enterprise, design for coordination: appoint a lead or shared orchestrator, log inter-agent messages and require read-confirmation before irreversible actions. Use benchmarks such as MAMUBench in procurement pilots rather than relying on single-agent demos.