Self-organizing agent teams average 66.7% vs 48.8% for best single model, Stanford and Together AI paper finds
james_y_zou · x · 2026-09-24
- A paper from Stanford and Together AI shows letting agent teams learn their own collaboration strategy far outperforms expectations.
- A self-organizing team of three models (o3-mini, Claude Sonnet 4, DeepSeek-V3) averaged 66.7% across five math and physics benchmarks.
- Comparisons: the strongest member alone scored 48.8%, and even a perfect router over members' independent answers only hit 59.0%. On AIME 2026 the team reached 71.2%, 13.4 points above the router.
- Method: one member reviews the team's earlier exchanges and rewrites the teamwork strategy — roles, phase ordering, participation, and how partial answers are merged.
- The authors note research on effective agent collaboration/communication is lacking and self-organizing agents are more capable than previously thought.
More from coding & agent
- User automates entire video clipping workflow with Nous Research's Hermes agent — Teknium · 2026-11-12
- As models improve, writing good evals gets harder and more important — amankhan · 2026-11-12
- Free AI game jam in SF Nov 14–15 with teams from Anthropic, Google, Meshy, OpenAI — chrisfirst · 2026-11-12
- Omarchy 4.6 lets you share local AI models across your tailnet, says DHH — alexcovo_eth · 2026-11-12
- Coding harnesses hand shell access to inference providers — 4-step gateway mitigation — shensi · 2026-11-12
- Slack's David Hoang: agents beat people on coordination, proposes 'boundary agents' — davidhoang · 2026-11-12