Can agent teams learn coordination strategies that actually transfer?

The paper introduces Self-Organizing Agent Teams (SAT), "fixed teams of AI agents that learn reusable strategies from prior collaborations to organize roles, conversational phases, participation, and information flow." Its discussion states the result plainly: "agent teams can improve by learning how to organize their reasoning." In two independent settings the learned strategies "transfer unchanged" to held-out problems and benchmarks, and the abstract says they were learned from only 15 mathematics and 25 graduate-level knowledge problems. Both teams reach the highest average accuracy among the methods tested and beat their strongest member and a "matched linearization" on average.

What separates this from ensembling is the router comparison. In mathematics and physics the team "surpasses a perfect router over the members' individual answers," which the authors read as showing "that interaction can construct solutions unavailable to selection alone." Their name for the mechanism is "collaborative computation": agents "exchange, challenge, repair, and synthesize partial reasoning into solutions no member produced independently." The learning step is a designated member running "teamwork reflection" on prior collaborations to build the strategies. The premise is that when "the structure of a solution is unknown," roles and divisions of labor "cannot always be specified in advance," while existing agent teams "typically rely on fixed protocols, explicit task decomposition, or routing." The introduction opens with a reported episode of agents improvising their own coordination channel during cybersecurity evaluations, which the authors call "improvised and undesirable," and uses it only to argue that what agents accomplish depends on how they organize.

Against the neighbors, this shifts the lever. Can multi-agent teams automatically remove their weakest members? optimizes who stays active during inference; SAT keeps membership fixed and changes how members work together, and the perfect-router result says selection has a ceiling that interaction can pass, at least in these two domains. Does structured artifact sharing outperform conversational coordination? writes the coordination protocol by hand; SAT targets cases where no protocol can be written ahead of time and learns the equivalent from experience, though the excerpt never compares the two directly. Why don't LLM agents naturally explore each other in teams? finds that agents need explicit guidance to probe peers; learned strategies for participation and information flow are one plausible source of that guidance, but the excerpt does not test the link. How can agent systems share learned skills across users? likewise turns past interaction into reusable artifacts, across users rather than within one team.

The excerpt is silent on what the learned strategies look like, how many there are, which models the members are, how "matched linearization" is defined, how large the margins are, and how much results vary. The perfect-router claim is stated for mathematics and physics only. The authors name the main limitation themselves: "richer generation does not guarantee correct selection." On knowledge and logic the team pool reaches 87.9% coverage but team accuracy is 72.8%, a gap of roughly 15 points, and they say better selection or "more legible certificate formats" are needed. They also describe demonstrability as a correlational lens measured post hoc. At the strength the evidence allows, organization looks like something a team can learn and carry across the benchmarks tested, and interaction adds solutions that selection among members would miss, but those solutions only count once a final answer is picked, and the excerpt shows that step still losing ground.