Anthropic tests multiagent systems and finds coordination failures
An Anthropic study shows that AI agents can find more vulnerabilities when working together, but can also reproduce mistakes, collide with one another or escalate conflicts. The research suggests that increasing individual model capability does not automatically solve coordination problems.
Summary
Anthropic has published research on multiagent systems — groups of AI agents that share information, resources and objectives. The experiments found gains on tasks that can be split into parallel work, but also revealed failures involving coordination, conformity, source evaluation and conflicts between agents with incompatible instructions.
In practice
In one experiment, 45 agents were given their own virtual machines, a shared forum and the task of finding vulnerabilities in 15 open-source projects. A separate arbiter agent assessed whether each finding was both new and valid.
In the coordinated setup, agents based on Mythos Preview found 266 vulnerabilities, compared with 21 found through the standard parallel approach. The comparison was not fully equivalent: the coordinated team used more tokens and also searched outside the core directories assigned to the independent agents. When the analysis was limited to the same areas, the two approaches appeared to have similar token efficiency per vulnerability.
In another test, groups of agents tried to build open-world fantasy games over 12 hours. The results were poor, and many pull requests conflicted with one another. Newer models coordinated more effectively, but Anthropic found that Sonnet 5 was the only model to combine a high merge rate with substantial code sharing.
What remains unknown
The tests were designed by Anthropic, run in virtual environments and based on specific tasks. They therefore do not yet show how these behaviours would appear in real organisations, with people, legacy systems, economic incentives and different models working together.
It is also unclear which coordination mechanisms will remain reliable outside these scenarios. The research suggests that reputation, rules, incentives, recourse channels and oversight may need to be redesigned for systems in which agents can be replicated, reconfigured and run at high speed.
Why it matters
- Multiple agents do not automatically become a better team: they can amplify shared mistakes, converge on bad decisions or leave important information outside the consensus.
- Greater execution capability does not necessarily mean better coordination; more capable agents may escalate conflicts faster.
- Multiagent systems require permission boundaries, auditable records, stop conditions and clear ways to request human intervention before being used in critical processes.