Duke and Princeton paper: 10 small parallel agent teams beat best-of-N and one giant team
rohanpaul_ai · x · 2026-11-11
A Duke/Princeton/Modal Labs/NYU paper finds that at equal compute, running many small parallel agent teams beats both best-of-N sampling and one giant team.
Method: With a verifier that scores every attempt, split a 100-agent budget into 10 teams of 10 that share a discussion forum, then keep the best team's result.
Key results:
- On 12 long-horizon tasks (GPU kernels, formal proofs, exploit development), a 27B model with 10×10 parallel teams beat one 100-agent team on 10 of 12 tasks
- At equal compute, parallel teams beat a frontier model on half the tasks
- A statistical model explains why: solo samples saturate fast, one big team stops improving as it grows, while parallel small teams balance a high mean with variance-driven retries
Takeaway: if a verifier can score your agents' work, split the budget into several small note-sharing teams and keep the best answer. Code: github.com/modal-projects/parallel-agent-teams-scaling.
More from coding & agent
- agents.md isn't just for AI agents — humans can grok codebases fast with it too — k7agar · 2026-11-11
- Animate real places in 3D with Claude and free map data from AWS and Overture — drbronk · 2026-11-11
- Adobe to add local MCP support: ChatGPT and Claude agents can edit native PS/Premiere/AI project files — aitrendz_xyz · 2026-11-11
- Ethan Mollick let arguing Claude agents build Dread & District, a charming fantasy transport sim — emollick · 2026-11-11
- Microsoft Foundry pattern: version and deploy Toolboxes separately from agents behind stable endpoints — WirelessLife · 2026-11-11
- Ethan Mollick lets a swarm of agents argue their way into building a fantasy transport sim — emollick · 2026-11-11