Two developers ask an AI coding assistant for changes in the same repository, neither aware of the other. Each edit may work alone but clash when combined.
Researchers built a test set of paired requests. Some need both changes to share an interface, a fixed way of calling each other's code; others demand incompatible edits to the same lines. They ran pairs of agents, programs driven by large language models (LLMs), and compared them with one agent given both requests at once.
The single agent won far more often: 62% versus 16% on conflicting pairs. Letting the two agents message each other helped only slightly, and cost more to run.