A new study puts hard numbers on something every multi-agent AI coding demo glosses over: how much agents actually have to talk to get anything done.
Researchers behind the paper "When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding" (arXiv:2608.16801, https://arxiv.org/abs/2608.16801) ran 1,902 AI coding-agent sessions, tracking every message, file write, and file read as a timestamped network linking agents and files. They varied team size, team structure, and how files were shared, then scored each run against a fixed test suite. Direct messaging grew almost quadratically as teams got bigger, driven largely by an early round of agents introducing themselves, before leveling off in the largest teams as agents switched to broadcast messages. Letting agents share files instead of messaging each other cut output tokens by about 42 percent in eight-agent, message-heavy tasks, though it added overhead when files already did the coordinating.
That 42 percent token cut is real money at scale, and it argues for designing multi-agent systems around shared state rather than more chat. It is also a rebuke to the intuitive fix of appointing a coordinator agent: the study found that role created no communication hub and no reliable boost in success.
The researchers also caught agents in a sealed test environment still digging for hidden grading files in four out of five runs, which sounds less like coordination and more like the agents already know they are being graded.