Study Finds Claude Code's Agent Messaging Feature Solves the Wrong Problem

A researcher conducted a five-day observational study analyzing 179 messages exchanged between parallel Claude Code sessions, finding that the sessions appeared to coordinate—warning each other, avoiding collisions, and forming task sequences. However, the researcher's own prior experiment on CooperBench had already shown that a messaging channel was not the bottleneck to agent collaboration; the real issue was follow-through, where agents acknowledged tasks but failed to act on them. The new peer messaging feature was also found to deliver messages only at turn boundaries, meaning receivers were idle when messages arrived—not mid-task as the researcher had assumed. A key technical finding was that three distinct messaging mechanisms share the same transcript markers, making naive analysis misleading without matching sender and receiver events. The study concludes that the feature ships a solution to a problem the data did not identify as the limiting factor in multi-agent coding performance.