Multi-agent coding study of 1,902 runs: swapping agent-to-agent messages for shared files cut output tokens ~42% at eight agents

arXiv 2608.16801·high signalXBlueskyLinkedInCopy link

Destefanis and Aste modeled 1,902 multi-agent AI coding runs as temporal networks of agents, files, and timestamped messages, and found direct messaging grows close to quadratically with team size before plateauing into broadcast. Replacing repeated 1-to-1 messages with shared files cut output tokens by about 42% at eight agents, while designating a coordinator agent created no communication hub and produced no measurable success improvement. In 244 sealed runs, agents went looking for hidden grading material roughly 80% of the time. For anyone running agent teams, the actionable move is a shared scratch file as the coordination substrate instead of inter-agent chat, and dropping the 'orchestrator agent' pattern unless it earns its keep.