Researchers tested whether a fixed budget of AI agents goes further when spent on several small teams. In their preprint, ten teams of ten agents outperformed 100 independent agents on 10 of 12 tasks. The comparison held the underlying 27-billion-parameter model fixed, which helps separate the way agents are organized from the choice of model. The preprint was submitted to arXiv on November 6, 2026.
How the teams use extra compute
Best-of-N sampling gives every agent a separate attempt and uses a verifier to pick the best answer. A single large team takes the opposite approach: it puts the entire budget into collaboration on one result. The paper's third approach asks small teams to collaborate internally while running the teams in parallel. Each team submits an answer, and the verifier chooses among those answers.
The two levels of organization matter for different reasons. The authors argue that collaboration improves the quality of a typical attempt, while repeated teams still give the verifier a chance to select an unusually good result. Putting every agent in one group gives up that second opportunity.
Where the gains appeared
The task set included GPU kernel optimization, formal proofs and exploit development. On Carleson Formalization, the authors report that the best team result scored 54.8 out of 100, compared with 19.0 for the best independent agent at the same 100-agent budget. Small teams also beat a single team of 100 on all four tasks where that comparison was run. The approach did lose to independent agents on Kimi-Linear Decode and, narrowly, on Active Learning.
What the experiment leaves open
The tasks all offered a way to score each attempt, and the study tested only four tasks with a 100-agent team. The findings are most useful when you can afford several attempts and have a verifier that distinguishes working results from plausible ones. For a workflow where judging an answer is itself the hard part, the results do not establish that team size will help.
Comments