Many small agent teams beat one big swarm, and the math says why
Say you have a budget of 100 agents and a verifier that can score every attempt. Do you run 100 solo agents and keep the best? Do you put all 100 on one team? A new paper says neither: run ten teams of ten in parallel and keep the best team's result.
That is the method in "Test-Time Compute Scales Better with Parallel Agent Teams" (arXiv 2611.09798, Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala, Zhuang Liu). According to the abstract, they tested it on 12 long-horizon tasks spanning GPU kernels, formal proofs and exploit development, with a 27B model. Parallel teams of ten beat parallel solo agents on 10 of 12 tasks, and beat one 100-agent team on all four tasks where they ran a team that large. At equal dollar cost, the same parallel teams beat a frontier model on half the tasks.
The explanation is the part to keep. Best-of-N with solo agents saturates because each solo attempt has a low average. One giant team scores higher on average, but it is a single draw, so you get no benefit from variance, no lucky tail. Many small teams give you both: a decent average per team, and enough draws for one of them to land high. It is the same logic as running several startups instead of one huge company or a thousand freelancers.
This lands on the same day Anthropic bragged about 1,000 Fable subagents building an OS. The paper suggests the shape of the swarm matters as much as its size, and that a mid-size open model arranged well can match a frontier model on some tasks for the same money. The big condition: you need a verifier that can score each attempt, which kernels and proofs have and most office work does not. Numbers here are from the abstract. The code repo linked from the paper, github.com/modal-projects/parallel-agent-teams-scaling, returned 404 when checked.
Source: https://arxiv.org/abs/2611.09798
← Back to all articles
That is the method in "Test-Time Compute Scales Better with Parallel Agent Teams" (arXiv 2611.09798, Junlin Wang, Haozhe Chen, Ang Li, Rahul Chalamala, Zhuang Liu). According to the abstract, they tested it on 12 long-horizon tasks spanning GPU kernels, formal proofs and exploit development, with a 27B model. Parallel teams of ten beat parallel solo agents on 10 of 12 tasks, and beat one 100-agent team on all four tasks where they ran a team that large. At equal dollar cost, the same parallel teams beat a frontier model on half the tasks.
The explanation is the part to keep. Best-of-N with solo agents saturates because each solo attempt has a low average. One giant team scores higher on average, but it is a single draw, so you get no benefit from variance, no lucky tail. Many small teams give you both: a decent average per team, and enough draws for one of them to land high. It is the same logic as running several startups instead of one huge company or a thousand freelancers.
This lands on the same day Anthropic bragged about 1,000 Fable subagents building an OS. The paper suggests the shape of the swarm matters as much as its size, and that a mid-size open model arranged well can match a frontier model on some tasks for the same money. The big condition: you need a verifier that can score each attempt, which kernels and proofs have and most office work does not. Numbers here are from the abstract. The code repo linked from the paper, github.com/modal-projects/parallel-agent-teams-scaling, returned 404 when checked.
Source: https://arxiv.org/abs/2611.09798
Comments