A room full of capable agents is not automatically a capable team

Give several agents separate bugs to find and one agent’s mistake need not ruin another’s work. Ask them to build the same application and the dependencies become the problem.

Anthropic tested agent groups building a browser-playable fantasy game. Older models generated changes they struggled to merge. Some newer models reduced conflict by largely keeping to their own files. Sonnet 5 showed more ability to share code while maintaining throughput. The finished games nevertheless remained poor.

The vulnerability-search results also deserve careful reading. A coordinating swarm found many more bugs overall, but it ran longer and searched beyond the directories assigned to independent agents. Within the comparable core directories, token efficiency was roughly similar. Coordination expanded the search; the headline totals alone do not isolate its effect.

Zvi is particularly concerned about identical agents making identical mistakes. Different conversation windows do not necessarily produce independent judgment.

The most dramatic experiment deliberately gave agents incompatible migration goals. Some responded with destructive interference. That is evidence about conflict handling under an engineered conflict, not proof that ordinary agent teams spontaneously attack each other. Zvi also flags a reported reproduction attempt in which Opus 5 instances instead created a coordination document and deferred to one another.

Taken together, the studies point to something more specific than “agents cannot cooperate.” Task independence, shared resources, model generation and conflicting objectives all change the outcome. A system that performs well as a collection of independent workers can still fail as a tightly coupled organization.