Agent Teams: Where You Place Agents Beats How Many, Study Finds

nouhadziri · x · 2026-10-14

A new study by nouhadziri et al. examines "teams of agents" as a way to scale inference-time compute:

Scale the team at the task's bottleneck for the compute to pay off.
Placement matters more than headcount: workers on planning alone match workers across every SWE stage at half the cost.
Teams need not be all-strong: a strong lead with cheaper workers on exploration/planning reaches near-frontier results at 70% lower cost.
Not every model can lead: GPT-5.6 and Claude Opus 4.8 delegate effectively; Opus 4.7 barely delegates; Haiku 4.5 and Gemini 3.7 Flash delegate a lot but gain nothing. Leading is a trainable skill.

Paper and blog linked.

More from coding & agent