Scaling agent teams: reach the frontier at 30% of the cost

This post introduces a controlled testbed for evaluating multi-agent system organization for coding tasks. It presents findings on how to optimize agent teams for effectiveness and efficiency by varying agent roles (homogeneous vs. heterogeneous), lead behavior (blocking vs. asynchronous), delegation strategy (minimal vs. PIV prompt, single-stage delegation), and team size. Key metrics like reward and cost are analyzed, showing that strategic delegation, particularly for planning, can achieve comparable results to full delegation at a fraction of the cost.