Researchers introduced Self-Organizing Agent Teams for AI agents. These teams learn reusable strategies from past work to organize roles and reasoning. The approach enables collaborative computation where agents exchange and synthesize partial reasoning. Teams averaged 66.7% accuracy across five benchmarks. This outperformed the strongest single agent at 48.8%. On AIME 2026, teams beat a perfect router by 13.4 points. Gains were largest when correct reasoning could be recognized. The study used 15 math and 25 graduate-level problems to train strategies. Strategies transferred unchanged to unseen benchmarks.
What it unlocks
AI agent teams can now learn reusable strategies to organize roles and reasoning without fixed protocols.
- 66.7% team accuracy
- 48.8% strongest member accuracy
- 13.4 points above perfect router on AIME 2026
- Spearman ρ=0.90, p=0.005
What you need to act on it
- fixed teams of AI agents
- prior collaborations to learn from
Sources