Measuring Coordination in Multi-Agent AI Coding Teams

A new arXiv preprint (2608.16801) introduces an instrument to measure coordination among AI coding agents. The study represents each run as a temporal network with agents and files as nodes, and timestamped directed edges for messages, file writes, and reads, each with an associated cost. Applied to 1902 runs across varying team sizes, structures, and file policies, the analysis reveals that direct messaging initially increases nearly quadratically with the number of agents, driven largely by an early round of introductions, but levels off as teams grow further. The work addresses a gap in current evaluations, which typically report only task completion and cost, leaving coordination unmeasured.