Researchers from Duke University, Modal Labs, Princeton University and New York University report that, with a fixed budget of 100 AI agents, splitting them into ten teams of ten and keeping the best team’s result scored 27% higher, averaged over 12 tasks, than running 100 agents alone and keeping the best one.
On one maths task, Carleson Formalization, the paper says the best of 100 solo agents scored 19.0 out of 100 while the best team scored 54.8, which is 2.9x, by our arithmetic. The paper, Test-Time Compute Scales Better with Parallel Agent Teams, was submitted to arXiv on 6 November.
Business Pill · Multi-Agent Systems
A one-minute explainer of multi-agent systems: how several AI agents split and share work. It teaches the general idea only and does not discuss this paper.
The key insight: A Duke, Modal Labs, Princeton and NYU preprint reports that ten teams of ten agents scored 27% higher than 100 solo agents, averaged over 12 tasks a verifier can score. On one maths task the best team scored 54.8 against 19.0, 2.9x by our arithmetic.
The authors compare three ways to spend the same agents when a verifier can score every attempt: best-of-N over individual agents, one big team, and what they call parallel agent teams. In their words, “Best-of-N sampling runs many independent agents and keeps the best result but saturates quickly”, while “One large team of collaborating agents can score higher but stops improving as it grows”.
Each team shares a forum. The paper says “Each agent has its own sandbox and context but shares a forum where any agent can post and read notes at any time”, with “no roles, no manager and no assigned work”. Agents ran in the Codex harness inside Modal sandboxes, and the main model was Qwen3.8-27B, according to the paper.
Best score on the Carleson Formalization task with a budget of 100 Qwen3.8-27B agents, from the 6 November preprint. The 2.9x ratio is our arithmetic.
What Was Posted on X
New Duke and Princeton paper finds that, at equal compute, many small agent teams in parallel beat both solo-agent best-of-N and one giant team.
With the same budget of 100 AI agents, running 10 small teams of 10 agents in parallel and keeping the best teamโs result beat both 100 solo agents and one 100-agent team.
With a verifier, teams usually run many solo agents and keep the best, but gains flatten fast. One large team does better at first, then stops improving as it grows.
They split the agent budget into teams of 10 that share a forum, then keep the best team’s result. Teamwork raises each attempt’s average, and several teams still give many tries.
If a verifier can score your agents’ work, split the budget into several small teams that share notes, and keep the best team’s answer.
Rohan Paul posted at 20:02 UTC on 10 November a summary of the Duke and Princeton paper: with the same budget of 100 AI agents, ten small teams of ten kept the best team’s result and came out ahead of both 100 solo agents and one 100-agent team. This piece takes its figures from the paper itself.
The Results the Authors Report
The abstract says “parallel teams of ten beat parallel agents on 10 of 12 tasks and beat one team of 100 on all four tasks where we ran such large teams”. The paper puts the margin over one team of 100 at 8% to 30%, and says parallel agents came out ahead only on Kimi-Linear Decode and, by a small margin, on Active Learning.
On cost, the paper says a Qwen3.8-27B team of ten “costs about as much as one or two GPT-5.6-sol agents”, and that at equal dollar cost the 27B teams scored higher than the frontier model GPT-5.6-sol on 4 of 8 tasks. Rohan Paul summarised the paper on X on 10 November.
Why It Matters for the Business
The authors argue that individual agents lose because their many samples have a low mean, and one large team loses because a single sample gets no gain from variance. The comparison lands as labs ship multi-agent features: we covered Google’s Antigravity /swarm command and a benchmark of models running agent teams this week, and a map of 67 firms selling agent deployment.
The Structural Read
As we read it, the paper treats how agents are grouped as a cost lever in its own right: the same 100 agents score differently depending on whether they work alone, in one crowd or in small teams. The authors limit the finding to tasks where a verifier can score every attempt.
Wang, Chen, Li, Chalamala and Liu, arXiv 2611.09798, 6 November 2026
“Sampling many small teams in parallel balances a high mean with the gain from variance”
Three Implications
FOR AGENT BUILDERS The authors say parallel teams of ten scored higher than parallel agents on 10 of 12 tasks at a budget of 100 agents.
FOR MODEL BUYERS The paper says a Qwen3.8-27B team of ten costs about as much as one or two GPT-5.6-sol agents, and the teams scored higher on 4 of 8 tasks at equal cost.
FOR COMPUTE PROVIDERS Three of the five authors list Modal Labs, and the agents ran inside Modal sandboxes, according to the paper.
The framework reads: “The next wave of AI diffusion depends more on organizational transformation than on model improvements.”
As we read it, the preprint is a small-scale test of that idea for agents themselves: the model stays fixed and only the grouping changes. The framework predates the paper and does not discuss it.
What Is Not Established
The results are the authors’ own and cover 12 tasks where a verifier can score every attempt; the paper does not test open-ended work without a scorer. It is a preprint, and we found no peer review or independent replication. Three of the five authors list Modal Labs, whose sandboxes the agents ran in.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
A Duke, Modal Labs, Princeton and NYU preprint reports that ten teams of ten agents scored 27% higher than 100 solo agents, averaged over 12 verifiable tasks, and 2.9x higher on one maths task, by our arithmetic. The authors say the 27B-model teams also outscored GPT-5.6-sol on 4 of 8 tasks at equal cost.
A note on sourcing. The findings, models and caveats are from the authors’ arXiv preprint of 6 November 2026, read on 10 November. The 2.9x ratio is ours. We did not contact the authors.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
