Researchers introduce Self-Organizing Agent Teams (SAT), fixed groups of AI agents that learn reusable collaboration strategies from as few as 15-25 prior problems rather than following fixed protocols or routing. Across five math and physics benchmarks the learned strategies transfer unchanged to unseen tasks, averaging 66.7% accuracy versus 48.8% for the best individual agent and 59.0% for a perfect-router baseline, with a 13.4-point gain on AIME 2026. The authors identify demonstrability, whether a team can recognize correct reasoning once produced, as the factor predicting when self-organizing collaboration helps.
