Skip to content- Sources: primary
- Summary: On ProgramBench, Nouha Dziri reports that agent teams beat a solo agent by up to about 10 points but cost 4 to 7 times as much, and that adding workers paid off only at the stage that bottlenecks the task, planning there and implementation on Commit0. A strong lead with cheap workers kept most of the reward while cheap leads capped results, and 84 percent of messages went from workers to the lead, with peer messaging adding no reward. A blocking lead outperformed an asynchronous lead under a minimal prompt because the async lead moved on before worker results arrived.
- Why it matters: Engineers building multi-agent coding harnesses should spend on the lead model and on parallel discovery at the bottleneck stage rather than spreading workers across every stage.