Moving Beyond Fixed AI Workflows

Most current multi-agent systems rely on rigid, pre-defined protocols, such as explicit task decomposition or static routing. These approaches fail when the structure of a solution is unknown. The authors introduce Self-Organizing Agent Teams (SAT), a framework where agents learn reusable strategies from prior collaborations to dynamically manage roles, conversational phases, and information flow. This allows for "collaborative computation," where agents actively exchange, challenge, repair, and synthesize partial reasoning.

Performance and Demonstrability

SAT demonstrates significant performance gains over traditional methods. Across five mathematics and physics benchmarks, these self-organizing teams achieved an average accuracy of 66.7%, compared to 48.8% for the strongest individual member and 59.0% for a perfect router. Notably, on the AIME 2026 benchmark, SAT outperformed the router by 13.4 percentage points.

The researchers identified that the effectiveness of these teams is tied to the concept of "demonstrability"—the ability of a team to recognize correct reasoning once it is produced. Across eight benchmarks, there was a strong correlation (Spearman ρ=0.90) between a task's demonstrability and the team's ability to outperform its strongest individual member. This suggests that organization is not just a management layer, but a core capability that allows AI teams to synthesize solutions that no single member could reach independently.