Illustrative image: a neural-network diagram. The paper's agent teams ran on a 27-billion-parameter open model. Image: Wikimedia Commons
Key Highlights
- The result: ten teams of ten AI agents scored 27% higher on average than 100 solo agents, given the same total budget of 100 agents.
- It also beat the big-team approach: parallel teams outscored one team of 100 agents by 8% to 30% on every task where a 100-agent team was tested.
- Small models, big effect: at equal dollar cost, teams built on a 27-billion-parameter open model beat a frontier model on 4 of 8 tasks.
- The catch: it only works on tasks where a verifier can score every attempt — code, maths proofs and similar checkable work — and the paper is a preprint, not yet peer-reviewed.
What happened?
On 6 November 2026, researchers from Duke University, Modal Labs, Princeton University and New York University posted a paper to arXiv titled "Test-Time Compute Scales Better with Parallel Agent Teams". The authors are Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala and Zhuang Liu. The paper asks a practical question that anyone building with AI agents faces: if you have a fixed budget of agents, how should you organise them?
There are two obvious answers, and the paper shows both are wrong. Run 100 agents independently and keep the best result (best-of-N sampling), and the scores level off quickly. Put all 100 in one big collaborating team, and the team stops improving as it grows. The researchers' alternative is simple: split the 100 agents into ten small teams of ten, let each team work on its own, and keep the best team's result. They tested it on 12 long-horizon tasks — writing GPU kernels, formalising mathematical proofs, and exploit development among them — using the open model Qwen3.8-27B in the Codex harness inside Modal sandboxes.
How parallel agent teams scale test-time compute
Test-time compute means spending extra computation while the model is answering, rather than while it is being trained: generate several candidate answers, check them, and keep the best. The paper's insight is about the shape of that spending. Individual agents produce many attempts, but each attempt is weak, so the average quality is low. One large team produces strong work, but it is a single attempt, so there is no second chance if it goes wrong. Ten small teams get both: each team is strong enough to do good work, and having ten of them means the best one can still win.
The numbers bear that out. At a total budget of 10 agents, parallel teams scored 10% higher than solo agents; at a budget of 100, the gap grew to 27%. On one maths task, Carleson Formalization, the best of 100 solo agents scored 19.0 out of 100 while the best team of ten scored 54.8 — nearly three times as much. The method won on 10 of the 12 tasks; solo agents won only on two, Kimi-Linear Decode and Active Learning, and by a small margin there.
| How to spend 100 agents | What happens | Result vs solo agents |
|---|---|---|
| 100 solo agents (best-of-N) | Many weak attempts; scores saturate quickly | Baseline |
| One team of 100 | Strong work, but a single attempt with no fallback | 8–30% below parallel teams |
| 10 teams of 10 (parallel agent teams) | Strong teams plus ten chances to get it right | 27% higher on average |
Illustrative image: a neural network. Parallel teams combine the strengths of solo sampling and team collaboration. Image: Wikimedia Commons
Why it matters for India readers
The most useful part of the result is the cost angle. The teams ran on a 27-billion-parameter open model, not a frontier system — and at equal dollar cost they still beat a frontier model (GPT-5.6-sol) on 4 of 8 tasks. For Indian developers and startups building coding assistants or support agents, that is practical: better organisation of the agents you already run can buy more accuracy than paying for a bigger model. It fits a pattern this blog has covered before — small, cheap tricks beating brute force, like deleting a tiny fraction of a model's weights to match full fine-tuning, or the cheap fix that lifted small-model maths scores. How you spend compute matters as much as how much you spend.
Should you care?
Where it does not work
Three limits matter. First, every task in the study had a verifier that could score each attempt — code with tests, proofs a checker can verify, maths with checkable answers. Open-ended work such as writing has no such scorer. Second, it is a preprint: it has not been peer-reviewed, and no independent team has replicated it yet. Third, the advantage is an average — on two of the twelve tasks, solo agents still won. Treat the 27% figure as the paper's headline average across its own benchmark, not a guarantee for your workload.
FAQs
What is test-time compute in simple terms?
It is the computation a model uses while answering a question, rather than while being trained. Thinking longer, trying several answers and checking them all are forms of test-time compute — accuracy bought at inference time, not from a bigger model.
What are parallel agent teams?
A way of spending a fixed agent budget: instead of many solo agents or one giant team, you run several small teams (ten agents each in the paper) in parallel and keep the best team's result.
Why do 10 teams of 10 beat 100 solo agents?
Solo agents make many weak attempts; one big team makes one strong attempt with no fallback. Small parallel teams get both: strong teams, and ten chances to pick the best.
Does this work for every AI task?
No. It was only tested where an automatic verifier can score every attempt — code with unit tests, formal proofs. For open-ended work without a checker, it has not been shown to help.
Takeaway: Splitting 100 AI agents into ten teams of ten beat both 100 solo agents and one team of 100 — a 27% average gain on checkable tasks, from organisation alone. The idea is free to try, but only where your work can be scored automatically.
If you build with agents, try it on one task you can score automatically before reorganising your setup — and read the full paper on arXiv first.
Sources
- Test-Time Compute Scales Better with Parallel Agent Teams — arXiv:2611.09798 (abstract, submitted 6 Nov 2026)
- Full text of the paper — arXiv HTML version
- FourWeekMBA — Ten AI Agent Teams of 10 Score 27% Above 100 Solo Agents
- QelvanoTech — Stop Running 100 Solo AI Agents, 10 Teams of 10 Beat Them by 27%
Comments
Post a Comment