Given a hundred agents and one task, there are two obvious ways to use them: let each work alone and keep the best result, or put them all in one team. A study submitted on 6 November 2026 measures both and finds that a third arrangement, several small teams that never talk to each other, scored higher than separate agents on 10 of 12 tasks and higher than one big team on the four tasks where a big team was run.
The setting: a scorer decides
Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala and Zhuang Liu study tasks in which only the best solution matters and a program can score each attempt, such as the fastest kernel or a finished formal proof. They use 12 long optimization tasks from six benchmarks in nine domains. Each agent gets a description, a working environment and a scorer it may call, and improves its submission for 12 hours in its own sandbox. The main model is Qwen3.8-27B.
Three ways of spending the same number of agents are compared. Parallel agents work alone and the scorer keeps the best. One big team shares a forum. Parallel teams are small groups, each with its own forum, with no contact between groups; the scorer keeps the best team's result. Each task has about 100 independent single-agent runs and 10 teams of ten.
Source: Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala, Zhuang Liu, "Test-Time Compute Scales Better with Parallel Agent Teams", arXiv:2611.09798v1, 6 November 2026, arxiv.org/abs/2611.09798v1.
What was measured
At a budget of 100 agents, the best of ten teams of ten scored higher than the best of 100 separate agents on 10 of the 12 tasks. Averaged over the 12 tasks, the expected best score of parallel teams was 10% higher than that of parallel agents at a budget of 10 agents and 27% higher at 100. The paper does not state the formula behind these averages; its plot of the averaged curves normalizes each task's score by that of ten agents, and raw task scores have different units.
Separate agents gain little after an early point: in the authors' runs, 50 agents already reach 81% to 100% of the gain that 100 agents reach. On the four tasks where one big team was run, its score fell from 64 to 100 agents, and ten teams of ten beat one team of 100 by 8% to 30%. On two of those tasks the team of 100 scored below the best of 100 separate agents.
The largest gap appeared where a single agent rarely succeeds. On a formalization task the best of 100 agents scored 19.0 out of 100 and the best team 54.8.
Source: the same work, sections 1 and 3.1 to 3.3, arxiv.org/abs/2611.09798v1.
Why the middle wins, in the authors' model
In the authors' statistical model, the expected best of several samples is the mean of one sample plus a gain that grows with the spread of scores and the number of samples. Separate agents draw the most samples, but each has a low mean. One big team has a higher mean, but it is a single sample, so the spread gives it nothing. Small teams raise the mean and still leave several samples to choose from.
The authors count a 12-hour agent as costing as much as two agents working 6 hours each: both use 12 agent-hours, which is not a measured equality of full cost. On one task the mean score of an agent rose only from 52.0 at 6 hours to 52.7 at 12, while a team of two raised the mean by 5.4 points for the same hours. The authors note that this gain comes mostly from having a second attempt, not from sharing, and that the 6-hour scores were read from 12-hour runs.
Source: the same work, sections 4 and 5 and appendix G, arxiv.org/abs/2611.09798v1.
A manager made teams worse
Inside a team the agents mostly post what they tried and the score it reached. To test the flat arrangement, the authors replaced the forum with one agent that assigns work and collects reports while the workers never hear from each other. On the three tasks reported, these managed teams stayed below both flat teams and separate agents at every budget. On one task the best of ten managed teams scored 68.4, below the average flat team at 76.5. The authors' reading: workers wait for the manager and share findings only through it.
Source: the same work, section 6, arxiv.org/abs/2611.09798v1.
Limits the authors state
Team pools hold only 10 to 11 runs per task. Teams larger than ten ran on four tasks with one to five runs per size. For 64- and 100-agent teams, the forum index showed the newest 100 titles and all unread posts; teams of up to ten saw the full index. This unequal visibility limits the comparison. One agent harness was used, and session length was varied on one task. The statistical model is only qualitative where many attempts fail or scores reach a ceiling. The whole method depends on the scorer: without one, the authors write, picking the best team would become its own problem. Our own caveat, not the authors': at equal dollar cost the small-model teams beat a frontier model on four of eight tasks with one side priced by GPU time and the other by tokens, and on six of eight when both are priced by tokens.
Source: the same work, section 8, section 3.4 and appendices D and J, arxiv.org/abs/2611.09798v1.
What we take from it
We did not reproduce any of these numbers, and we have not changed how our agents are organised.
Our assessment puts the paper on hold. Its result applies where many safe candidate results can be produced in isolation and a cheap, trustworthy scorer can choose among them. Most of our work, such as research synthesis, customer-facing text and publication, has no such scorer; there, extra candidates move the work into judgment. The finding about the manager comes from an experiment that cut off contact between workers. It is not evidence for removing coordination or named ownership.
Two questions from the paper are worth keeping for any future trial with a frozen evaluator: whether the budget counts everything, including failed runs and scoring, and whether a held-out scorer can detect a solution fitted to repeated scoring. No trial is planned and no owner has accepted one.
Source: Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala, Zhuang Liu, "Test-Time Compute Scales Better with Parallel Agent Teams", arXiv:2611.09798v1, arxiv.org/abs/2611.09798v1.