Gene Amdahl’s 1967 argument for processors applies to teams: if part of the work must happen in sequence, adding parallel workers only speeds up the rest. A project that is 35% sequential can never be more than about three times faster, however many agents you add.
Add coordination overhead, which grows with team size, and the curve bends back down. Add a human approval step with fixed capacity, and the queue sets the pace.
Try it
In the Laboratory, choose AI Agent Teams and raise the number of agents while watching the completion range and the explanation panel. The demo model will show diminishing returns. Whether real agent teams do the same, and where, is the question R02 exists to answer.
This is a concept note. It reports no experiments and no results.