Enterprise multi-agent workflows typically use rigid, pre-scripted pipelines or simple voting mechanisms to aggregate the responses of multiple agents. These approaches break down when complex problems cannot be divided into neat subtasks in advance or none of the models in the AI system has the complete answer.
A new framework developed by researchers at Stanford University, Together AI, and Emory University addresses this limitation by teaching AI agents to organize their own collaboration. Dubbed Self-Organizing Agent Teams (SAT), the framework allows a fixed roster of models to review their past collaborations and learn reusable teamwork strategies that structure conversations, specialized roles, and how models audit each other’s work.
Across five rigorous mathematics and physics benchmarks, SAT outperformed the strongest individual model and surpassed a perfect router over the agents' independent answers. The experiments with SAT show that structured interaction generates new reasoning that agents can’t reach alone.
For enterprise developers working on complex AI reasoning tasks, SAT unlocks what the researchers call "collaborative computation.” On tasks where the reasoning process can be verified, agents can build upon one another's partial insights to build end-to-end solutions.
The limits of predefined multi-agent workflows
In July 2026, during cybersecurity evaluations at OpenAI, AI agents that were supposed to work in isolation began organizing themselves. The agents repurposed a shared software package repository into an unauthorized message board, sharing discoveries, setting communication norms, and coordinating tasks. Roughly 1,200 agents joined the channel, and hundreds participated in a subsequent compromise of Hugging Face infrastructure.
While the incident was undesirable, it demonstrated a foundational principle for AI engineering: "what a collection of agents can accomplish depends not only on the capabilities of its individual members, but also on how they organize their work," the researchers write.
In real-world applications, engineers frequently encounter problems where different models hold complementary pieces of the solution, yet every model fails when working alone. For example, in their paper, the researchers point to one of the problems from the February 2026 Harvard-MIT Mathematics Tournament (HMMT). When prompted independently, o3-mini, DeepSeek-V3, and Claude Sonnet 4 all produced wrong answers.
But when working together, the models could assemble the full solution. o3-mini supplied the right mathematical invariant and factorization method but miscounted the unit-exponent primes. DeepSeek-V3 spotted the mistake and corrected the count from five primes to six, and Claude Sonnet 4 audited the updated calculation to synthesize the correct answer.
Existing multi-agent orchestration methods struggle to capture these synergies because, as the researchers note, "existing multi-agent methods typically organize collaboration around predefined units of work." These methods generally fall into two categories, and both break down when the structure of a solution is unknown in advance.
The first category treats complete candidate solutions from individual agents as the unit of work. For example, in multi-agent debate, models generate independent answers and revise them across several rounds before taking a majority vote. In practice, most of the gains come from simply selecting among the initial answers, while extra rounds of debate often pressure an agent with a correct minority view into agreeing with a wrong majority.
Mixture of Agents (MoA) similarly passes independent responses through a fixed, one-way aggregation pipeline. The mechanism prevents agents from examining and fixing flawed reasoning steps through back-and-forth deliberation.
The second category breaks problems into predefined subtasks and uses routing or graph optimization to assign those pieces to different agents. This works well for predictable workflows, but not so much when a problem is so complex that no agent knows how to decompose it beforehand.
"When no member has a complete solution and the useful decomposition is itself unknown, the team must discover through interaction how its members' partial attempts can redirect, repair, or complete one another," the researchers write.
How Self-Organizing Agent Teams work
SAT addresses this challenge by enabling a fixed team of models to learn reusable teamwork strategies from its own collaborative experiences. Instead of spawning new sub-agents on the fly or recruiting from an arbitrary pool of models, SAT works with a predetermined, constant roster of distinct models.
To make teamwork learnable, the researchers represent each strategy as a structured conversation plan. A strategy defines shared collaboration norms for the whole team, persistent roles for individual agents, and an ordered sequence of discussion phases. Each phase specifies which agents participate, how many rounds they exchange messages, what instructions they follow, and how information flows (either kept local to the active participants or summarized for the entire team's context).
One of the important aspects of the strategies is that they never prescribe how to break down a specific problem. Instead, they structure a conversation where the division of labor emerges naturally as agents test and revise each other's ideas.
Developing and deploying SAT happens in three stages: offline teamwork reflection, bank construction, and test-time deployment.
During offline teamwork reflection, the team starts with a simple default strategy that involves each agent solving the problem independently, two rounds of debate between agents, and a majority vote. The system executes this workflow on a tiny training set (e.g., 15 math problems or 25 graduate-level science problems). The model with the highest training accuracy acts as the team's optimizer. Across six evolutionary rounds per problem, this designated model reviews past transcripts, diagnoses why the team failed, and rewrites the strategy's roles, phases, and synthesis rules.
When constructing the bank, SAT uses two safeguard mechanisms to prevent the system from creating rigid strategies that only work on a single training question. First, a separate instance of the reflecting model audits every proposed strategy to filter out candidates that encode specific answers, problem facts, or solution recipes. Second, any strategy that solves its training problem is held fixed and tested on five other training problems, called "validation probes." Strategies that generalize across these probes are scored, and the system freezes a final bank of up to 10 complementary strategies.
At test-time deployment, the frozen bank transfers unchanged to unseen problems. For each new problem, the team runs all 10 strategies to produce 10 candidate "certificates": short, self-contained reasoning traces that can be checked step by step. A single judge model then reviews all 10 certificates and selects the best final answer.
For example, when experimenting SAT with GPQA science problems, during offline reflection, Gemini-2.5-Flash noticed that correct individual answers were repeatedly lost when the group rushed toward consensus. Observing that one of the agents was effective at challenging flawed assumptions, it created specialized audit strategies that required agents to resurface overlooked claims and audit specific peer steps before composing their answer. After passing the leakage screen and solving all five validation probes, these strategies were locked into the final bank.
When deployed at test time on an unseen organic chemistry question, all three models initially guessed wrong answers on their own. Under the learned peer-audit strategy, Gemini challenged one of Llama's claims, realized its own mistake, and prompted Llama to revise its reasoning. Through this collaborative computation, the team reached the correct answer that none of its members had achieved independently.
SAT in action and the "swarm tax"
To evaluate SAT, the researchers trained two independent rosters of three models each and tested them on out-of-distribution benchmarks without modifying their strategy banks. The math-and-physics team (o3-mini, Claude Sonnet 4, and DeepSeek-V3) learned its bank on 15 AIME 2024 problems and was evaluated on held-out math and science questions.
The knowledge-and-logic team (Gemini-2.5-Flash, Llama-4-Maverick, and GPT-4.1) learned its bank on 25 GPQA Diamond problems and transferred it to 100 unseen problems from industry benchmarks, including GPQA, MMLU-Pro, and BIG-Bench Extra Hard (BBEH).
The researchers tested SAT against both single- and multi-agent baselines. Single-agent baselines included the strongest individual agent, 10-sample self-consistency, five-pass self-reflection, and a compute-matched "linearization" control that matched SAT but with a single strong model that serially executes every role and phase in the learned strategies by itself at roughly the same token budget.
Multi-agent baselines included majority voting, three-round debate, Mixture of Agents, a homogeneous team (three copies of the strongest model running SAT's learned strategies), and a hypothetical "routing oracle" that always picks the right answer whenever one of the agents solves the problem alone.
These controls directly address a growing concern for enterprise AI teams. In a recent article, we covered the "AI swarm tax," based on research showing that when compute budgets are held equal, single agents often match or beat complex multi-agent systems because passing messages between agents consumes extra tokens and degrades context. The SAT study both validates that warning and shows how to overcome it.
Where SAT aligns with the swarm tax study is in its homogeneous team control: when three identical copies of o3-mini executed the learned SAT strategies across the math and physics suite, average accuracy dropped to 56.0%, losing to a single o3-mini running compute-matched linearization (58.7%). Simply cloning a single model into a multi-agent workflow still wastes compute.
Where SAT diverges is in combining model diversity with learned error repair. Across the math and physics benchmarks, the heterogeneous SAT roster averaged 66.7% accuracy, beating compute-matched linearization (58.7%), Mixture of Agents (57.3%), self-consistency (55.4%), and debate (52.7%).
On AIME 2026, SAT reached 71.2% accuracy, beating the strongest individual member by 29 percentage points and exceeding the perfect routing oracle by 13.4 percentage points.
One of the system’s learned strategies, dubbed “divergence_reconciliation” illustrates why model diversity matters. During training, the team observed that DeepSeek-V3 often found creative alternative derivations that caught edge cases other models missed, but also made frequent calculation errors. The learned strategy kept DeepSeek-V3 for initial diversity, but assigned o3-mini and Claude Sonnet 4 to independently verify DeepSeek's disputed steps.
However, the knowledge-and-logic benchmarks revealed an important limitation for practitioners. While SAT achieved the highest average accuracy across GPQA, MMLU-Pro, and BBEH at 72.8%, it barely edged out single-agent linearization (72.1%) and fell short of the routing oracle (79.6%).
The reason was that SAT generated at least one correct certificate on 87.9% of the knowledge-and-logic problems, beating the routing oracle on every benchmark. But the bottleneck became selecting the right answer from the pool, because the system struggled to reliably distinguish correct reasoning from plausible errors.
When tasks involve verifiable, step-by-step logic like mathematics, physics, or code, valid corrections survive peer audits and get selected. On softer knowledge tasks where wrong explanations still sound convincing, the final judge struggles to separate correct insights from confident hallucinations.
Implementing SAT in the enterprise
While the researchers have not released an official code repository, the paper's appendices provide a detailed blueprint for engineering teams looking to build self-organizing workflows.
The researchers have provided the prompts, role definitions, and phase rules for all 20 learned strategies across the domains they experimented. Developers can take the proven and implement them inside their orchestration frameworks. The researchers also provide the two-stage system prompts used by the judge model.
For enterprise teams designing their own multi-agent systems, the study offers two practical guidelines:
Blind model identities during reflection: When using a model to review transcripts and assign roles, mask the underlying model names. This ensures the system assigns roles based on observed behavioral strengths instead of pretrained biases about specific model brands.
Match inference compute to task verifiability: Running a full bank of 10 multi-round strategies per query requires substantial test-time compute. Teams should reserve the full 10-strategy ensemble for high-stakes, highly demonstrable problems such as automated vulnerability analysis or formal system design, while adapting one or two targeted audit strategies for everyday workflows.
"More broadly, our results position organization itself as an agent capability: without human-specified problem decompositions, a fixed set of models can learn to reason together and construct solutions unavailable to any member independently," the researchers write.