Stanford's SAT Framework Teaches AI Teams to Outsmart Any Solo Model

Researchers at Stanford and Together AI have introduced Self-Organizing Agent Teams (SAT), a method for teaching a fixed roster of large language models reusable collaboration strategies. The teams learn from 15 to 25 scored problems, freeze their strategies, and apply them unchanged to new benchmarks. In the reported experiments, those conversations sometimes produced correct answers that no model had reached independently.

Why fixed workflows stall

Common multi-agent systems prescribe the collaboration pattern before seeing the task. That approach works when the designer already knows how information should move, but it can fail when the useful division of reasoning emerges only during the discussion.

• Debate repeatedly exposes agents to competing answers. Much of its measured gain may come from selecting a strong initial answer, while later rounds can suppress a correct minority view.
• Mixture of Agents passes proposals through a one-way pipeline to a synthesizing model.
• Workflow search routes predefined subtasks through a generated process, which assumes that the task decomposes cleanly.

SAT addresses cases where each model holds only part of the solution and the appropriate reasoning structure remains unknown. It searches for conversation patterns that let agents propose, challenge, repair, and verify one another’s work.

Teamwork becomes a small program

SAT represents each collaboration strategy in a compact domain-specific language. The representation controls who participates, what each model sees, how long discussion continues, and how the team reaches a final answer.