Duke and Princeton paper finds small parallel agent teams beat solo agents at equal compute

Duke and Princeton researchers report that, with a budget of 100 agents, 10 parallel teams of 10 agents, keeping the best result, outperformed both 100 solo agents and one 100-agent team.

The paper says a verifier can score each attempt, and that a single large team improves at first but stops gaining as it grows.

The authors describe small teams that share notes as a way to keep many attempts while raising the average quality of each. The report, posted on X by Rohan Paul, summarizes the paper and does not include benchmark names or figures.