Anthropic measures failure modes in multiagent systems

Anthropic reports experiments in which agents coordinated through shared forums and repositories. In one vulnerability-search test, a 45-agent swarm used peer review and an arbiter. In a separate software task, the study tracked merged pull requests and cross-agent code sharing. Why it matters: A team of agents can appear busy while leaving conflicting pull requests unmerged or avoiding shared work. These outcomes are visible in the workflow and can be measured before a multiagent setup is trusted with a larger codebase.

Try this: In one multiagent coding trial, record the share of pull requests merged. Check whether agents edit shared files or stay in separate ownership lanes.