Duke and Princeton Study Finds Small Parallel Agent Teams Outperform Large Ones

A new Duke and Princeton paper finds that, at equal compute, splitting agents into multiple small parallel teams and keeping the best team's result beats both solo-agent best-of-N and one giant team. With a budget of 100 agents, 10 teams of 10 agents running in parallel and selecting the best team outperformed 100 solo agents and a single 100-agent team. The authors recommend splitting budgets into small note-sharing teams and keeping the best answer when a verifier can score results.