Collective intelligence depends not only on what team members know but also on how they organize their work. When the structure of a solution is unknown, useful roles and divisions of labor cannot be predefined; teams must learn to organize reasoning from experience as it unfolds. Human teams adapt this way naturally, whereas existing AI agent teams rely on fixed protocols, explicit task decomposition, or routing.
We introduce Self-Organizing Agent Teams (SAT), fixed groups of AI agents that learn reusable strategies from prior collaborations. These strategies govern role assignment, conversational phases, participation, and information flow. With them, agents engage in collaborative computation: they exchange, challenge, repair, and synthesize partial reasoning into solutions no single member could produce independently.
In two independent settings, we learned teamwork strategies using only 15 mathematics and 25 graduate‑level knowledge problems, and transferred them unchanged to unseen benchmarks. Across five mathematics and physics benchmarks, SAT achieved an average accuracy of 66.7%, compared to 48.8% for its strongest member, 58.7% for compute‑matched inference by that agent, and 59.0% for a perfect router over members' independent answers; on AIME 2026 it outperformed this router by 13.4 points.
Because gains varied across benchmarks, we investigated when self‑organizing collaboration helps. Measuring demonstrability—the organizational‑psychology construct of whether a team can distinguish correct from incorrect reasoning—across eight benchmarks revealed a strong correlation with improvement over the strongest member (Spearman $\rho=0.90$, $p=0.005$). Teams benefit most when correct reasoning can be recognized as soon as it appears.
Broadly, these findings suggest that organization itself can become an agent capability: agent teams can learn how to reason together and produce solutions their members could not reach independently.
Review