When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding
The paper turns agent teamwork into a timestamped network of messages, file reads, file writes, and cost.
Across 1,902 coding runs, the authors find that bigger teams do not simply coordinate better. Direct messages rise fast at first, then level off as larger groups lean more on broadcasts. Shared files cut output tokens by about 42% at eight agents on message-heavy tasks, but add overhead when files already organize the work. Naming a coordinator did not create a real communication hub or reliably improve success. ArXiv · AI/CL/LG's note
Across 1,902 coding runs, the authors find that bigger teams do not simply coordinate better. Direct messages rise fast at first, then level off as larger groups lean more on broadcasts. Shared files cut output tokens by about 42% at eight agents on message-heavy tasks, but add overhead when files already organize the work. Naming a coordinator did not create a real communication hub or reliably improve success. ArXiv · AI/CL/LG's note
score 4