Robert, your critical question about coordination overhead eating parallel gains is the right one to ask. I tested this directly on launch day with 4 Opus 4.6 agents working simultaneously on different game components. In my case, the coordination cost was low because the filesystem-based task board kept communication lightweight. Each agent owned its domain and only coordinated at integration points. The 16-agent C compiler project is impressive but a clean-room scenario. My experiment was smaller scale but on a real project with messier requirements. The overhead was manageable when agent boundaries aligned with module boundaries. Full results here: https://thoughts.jock.pl/p/opus-4-6-agent-experiment-2026
Robert, your critical question about coordination overhead eating parallel gains is the right one to ask. I tested this directly on launch day with 4 Opus 4.6 agents working simultaneously on different game components. In my case, the coordination cost was low because the filesystem-based task board kept communication lightweight. Each agent owned its domain and only coordinated at integration points. The 16-agent C compiler project is impressive but a clean-room scenario. My experiment was smaller scale but on a real project with messier requirements. The overhead was manageable when agent boundaries aligned with module boundaries. Full results here: https://thoughts.jock.pl/p/opus-4-6-agent-experiment-2026
One of the reasons claude-mpm works is the PM understands agent authority and delegates consistently. With peer systems the risk is as you describe.