A new Stanford University paper introduces a framework in which groups of AI agents learn how to collaborate from a small set of past experiences, then apply those teamwork strategies to entirely new problems. The approach, called Self-Organizing Agent Teams (SAT), achieved an average accuracy of 66.7% across five math and physics benchmarks, comfortably outperforming the best individual agent in the group, which scored 48.8%.

How SAT works

Conventional multi-agent AI systems typically follow a rigid protocol: agents debate a problem, then vote on an answer. SAT departs from that template by allowing agent teams to develop their own organizational structures, including roles, participation rules, conversational phases, and patterns of information flow.

A key insight is that these strategies can be learned from remarkably small datasets. SAT teams derived their collaborative playbooks from just 15 AIME 2024 problems or 25 GPQA Diamond problems, then transferred those strategies unchanged to separate benchmarks they had never encountered. AIME, the American Invitational Mathematics Examination, is a competition math test that has become a standard yardstick for frontier-model reasoning, while GPQA Diamond is the hardest slice of a graduate-level science benchmark written to defeat quick web lookups. That a few dozen problems sufficed for strategies robust enough to transfer intact frames teamwork itself as a learnable, reusable capability rather than per-task engineering.

The concept draws heavily on organizational psychology. Agents exchange reasoning, challenge one another's logic, and combine partial solutions to reach answers that no single agent could produce alone.

The model lineup reads like an AI all-star roster, spanning systems from several labs in both closed and open-weight form. Math and physics tasks featured o3-mini, Claude Sonnet 4, and DeepSeek-V3, while knowledge and logic benchmarks used Gemini-2.5-Flash, Llama-4-Maverick, and GPT-4.1.

The numbers that matter

SAT's 66.7% average accuracy did not only surpass individual agents. It also outperformed compute-matched single-agent inferences, which reached 58.7%, and a routing oracle constructed from the agents' individual answers, which scored 59.0%. A routing oracle selects the best individual answer for each problem after the fact, so beating it indicates that the team is genuinely producing better reasoning rather than cherry-picking.

On AIME 2026 specifically, SAT reached 71.2% accuracy, exceeding the routing oracle by .4 percentage points.

One of the most telling findings involves what the researchers call "demonstrability"—essentially, how easily correct reasoning can be distinguished from incorrect reasoning once it surfaces in conversation. Performance gains correlated with demonstrability at a Spearman correlation of 0.90. When sound reasoning emerges, these teams recognize it and build on it. For practitioners, the correlation reads as a deployment signal: SAT's advantages concentrate where a team can check whether a given line of reasoning holds up.

Why this cuts against conventional wisdom

Related Stanford work published earlier in 2026 found that single agents often match or outperform multi-agent setups when given equal compute budgets. SAT directly addresses that critique. By focusing on learned organizational strategies rather than brute-force debate protocols, the framework demonstrates that collaboration can yield genuine advantages beyond what ensembling or individual inference provides. The gap between SAT's 66.7% and the compute-matched single agent's 58.7% suggests the value lies not in having more models, but in having models that know how to work together—a finding that positions orchestration design as a learnable discipline rather than a template to hand-tune.

The strong correlation with demonstrability also points to a natural boundary condition. SAT performs best on problems where correct reasoning is recognizable when it appears in conversation—domains with verifiable logic chains such as math, physics, and structured reasoning. How far the same playbook approach extends into open-ended work without such checks is the question the results leave open.