Parallel agent teams beat best-of-N and one big team at equal test-time compute
rohanpaul_ai · x · 2026-11-11
A Duke/Princeton paper proposes parallel agent teams: sample many small collaborating teams and keep the best team's result. With a 27B model and 100-agent budget across 12 long-horizon tasks (GPU kernels, formal proofs, exploit dev), ten 10-agent teams beat parallel solo agents on 10/12 tasks, beat one 100-agent team on all 4 such tasks, and matched a frontier model on half the tasks at equal dollar cost. A statistical model explains: solo agents lose on low sample mean, one large team loses because a single sample gains nothing from variance.
More from coding & agent
- Memgraph open-sources Context Graph: shared cross-session memory for Claude Code and Codex — JeremyCMorgan · 2026-11-11
- Fable 5.2-generated PR touches 3,000+ files, adds 441,926 lines in one change — cramforce · 2026-11-11
- Birthday gift: an AI agent emergently deployed the exact observability stack I wanted — sloppenheimer · 2026-11-11
- No Claude account? This DSH plugin lets you try Claude-style Whiteboards anyway — dotey · 2026-11-11
- Running 6 concurrent ds41-flash agents: persistent enough to need pi-durable coordination — dbreunig · 2026-11-11
- OpenAI roundup: architecture details of four GPT-Live voice agent builds — pbbakkum · 2026-11-11