Self-Organizing Agent Teams (SAT)

A fixed team of agents learns, from a small set of prior collaborations, a reusable strategy for organizing roles, conversational phases, participation, and information flow, then reapplies that learned strategy unchanged to new, unseen tasks.

Properties

family: Decomposition & Roles
maturity: frontier
diagram: roles
whyMultiAgent: Organization only exists across distinct agents that exchange, challenge, repair, and synthesize partial reasoning into solutions no single member could produce; there is no team structure to learn or reuse with only one model acting alone. Distinct from the already-tracked role-specialization pattern, which assigns personas and expertise to agents upfront by a designer: here the team learns its own division of labor and conversational structure from limited prior experience, and that learned strategy transfers unchanged to unseen problems, so the organization itself becomes a trained capability rather than a designer's choice. Also distinct from topology-task-routing, which classifies each incoming task into one of a few canonical orchestration shapes per query: SAT's team composition stays fixed, and the same learned strategy is reused regardless of the specific query.
mast: [Cat 1]
useCases: [Mathematics and graduate-level reasoning benchmarks solved by a fixed agent team][Collaborative computation on tasks where correct reasoning can be recognized once it appears (high demonstrability)][Reusable teamwork strategies that transfer across unseen problem sets without retraining]
source: Self-Organizing Agent Teams · arXiv:2609.22682

Provenance

Source: https://arxiv.org/abs/2609.22682

Properties

family

Decomposition & Roles

maturity

frontier

diagram

roles

whyMultiAgent

Organization only exists across distinct agents that exchange, challenge, repair, and synthesize partial reasoning into solutions no single member could produce; there is no team structure to learn or reuse with only one model acting alone. Distinct from the already-tracked role-specialization pattern, which assigns personas and expertise to agents upfront by a designer: here the team learns its own division of labor and conversational structure from limited prior experience, and that learned strategy transfers unchanged to unseen problems, so the organization itself becomes a trained capability rather than a designer's choice. Also distinct from topology-task-routing, which classifies each incoming task into one of a few canonical orchestration shapes per query: SAT's team composition stays fixed, and the same learned strategy is reused regardless of the specific query.

mast

[Cat 1]

useCases

[Mathematics and graduate-level reasoning benchmarks solved by a fixed agent team][Collaborative computation on tasks where correct reasoning can be recognized once it appears (high demonstrability)][Reusable teamwork strategies that transfer across unseen problem sets without retraining]

source

Self-Organizing Agent Teams · arXiv:2609.22682