Researchers at the University of Washington and the Allen Institute for AI say AI agent swarms can learn how to organise themselves instead of following a plan a developer writes. In a paper posted to arXiv on 5 November, they report that their method, s3, solved over 25 times more coordination problems than the strongest baseline when tested with twice as many agents as in training.
The test is small. The agents are copies of Qwen3 models with 0.6 billion and 1.7 billion parameters, on two research tasks, and the authors say each condition had one training run. The swarm’s cost also grows fast: if every agent broadcasts once, four agents make 12 message deliveries and sixteen make 240, by our arithmetic from the paper’s formula.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on the model and the harness. It teaches the general idea only and does not discuss the s3 paper.
The key insight: A University of Washington and Allen Institute for AI paper says agent swarms can learn their own coordination through cooperative reinforcement learning, with no predefined roles, and solved over 25 times more coordination problems than the strongest baseline at eight agents. The tests used small Qwen3 models on two research tasks.
Many multi-agent systems use an organisation their designers specify in advance, the paper says, such as a central orchestrator that splits the task or a leader that directs the others. In s3, a team is several copies of one model. No agent is a leader, no agent has a role, and there is no turn order.
Each agent can message any teammate or all of them and read its inbox when it chooses. All agents share one reward: the team succeeds if any agent reaches the correct answer, minus a charge for the share of the team’s token allowance it used.
The authors are Kunal Jha, Gonçalo Faria, Shuyue Stella Li, Scott Geng, Faeze Brahman, Noah A. Smith, Yulia Tsvetkov, Max Kleiman-Weiner and Natasha Jaques. The paper was highlighted on X on 9 November by elvis (@omarsar0), whose post had about 18,300 views.
Success of four-agent swarms on held-out problems in the paper’s information-aggregation task, by training method, from Table 7 of the s3 paper (mean of six seeds).
The main focus right now is training agent swarms to coordinate and cooperate better. Here is a promising new work on this.
The problem:
Multi-agent systems usually follow a plan that a developer writes ahead of time.
The plan decides which agent does what and when agents message each other, and it usually stays the same across tasks.
What they propose:
s3 has the agents learn that plan instead. Agents learn through cooperative RL when to work alone, when to share findings, and when to read teammates’ work. There are no predefined roles.
With matched training and inference budgets, s3 swarms tested at twice their training size solve over 25x more coordination problems than the strongest baseline. In parallel exploration, communicating agents repeat nearly 50% less work and beat best-of-N by up to 9.4 points.
The agents develop complementary roles on their own, and those roles carry over to bigger teams.
A post by elvis (@omarsar0) at 14:01 UTC on 9 November calling s3 a ‘must-read paper on scaling multi-agent swarms’ and summarising its claims: agents learn the plan through cooperative RL with no predefined roles, over 25x more coordination problems solved at twice the training size, and up to 9.4 points over best-of-N. Each figure matches the paper; the appendix shows the 9.4-point gain is not significant after the authors’ correction.
The Numbers in the Tables
In the first task, each of four agents holds one private number and the swarm must return the full sorted list. Trained with the shared reward, s3 reached 93.2% success on held-out problems. Agents trained to imitate successful swarms reached 34.9%, and agents each rewarded only for their own answer reached 12.5%.
The 25-times figure is from the paper’s abstract and refers to this task at eight agents. The main text says s3 then “succeeds more than twenty times as often as online SFT”, and the appendix puts the gap at 52.6 percentage points. The paper adds that all methods struggle at sixteen agents.
In the second task, agents search a tree of rooms for an exit. The room overlap of communicating agents was nearly half that of agents with messaging switched off, the paper says, though it notes the overlap measure is “not a pure measure of wasted effort”.
The headline 9.4-point gain over best-of-four, independent copies working alone, is significant before the authors’ statistical correction but not after, according to the appendix. At six agents, the gain is 7.3 points and holds after correction.
What It Costs to Run
The paper charges every swarm for tokens. In the search task, each agent gets 8,192 tokens, so a team’s allowance is 32,768 tokens at four agents and 131,072 at sixteen, by our arithmetic. One training run for that task used two Nvidia H200 GPUs for training and six for inference, the appendix says.
Messages are the other cost. The paper notes that each agent receives N−1 messages, which take up context space and leave less room for its own reasoning. By our arithmetic, going from four agents to sixteen multiplies the deliveries from one broadcast each by 20.
The trained swarm kept its token efficiency: the paper reports no statistically detectable decrease against best-of-N at six agents and above after correction. In the sorting task, s3 agents learned to send about one broadcast each, the paper says.
Why It Matters
Who writes the plan for a group of agents is a business question as much as a research one. In the systems the paper describes, the plan is a harness designed in advance; in s3 it is learned in training, inside the one policy every agent shares.
As we read it, the paper moves the coordination plan from a harness designed in advance into the trained policy. It also shows the cost side: message deliveries from one broadcast each rise from 12 to 240 as a swarm grows from four agents to sixteen, by our arithmetic.
Jha et al., s3 paper, arXiv 2611.08615
“Hand-designed organizations are fixed guesses about how a team should work, and they often break down”
Three Implications
FOR ORCHESTRATION In s3, no agent is a leader and there is no turn order; the shared team reward is what the agents are trained on.
FOR COMPUTE The paper charges each swarm for the share of its token allowance it uses; one search-task training run used two H200 GPUs for training and six for inference.
FOR THE EVIDENCE The 9.4-point gain over best-of-four is significant before the authors’ correction but not after; the 7.3-point gain at six agents holds after it.
The framework says: “Value is migrating upward from the model toward the harness.”
As we read it, s3 tests the opposite direction in a lab setting: coordination that a harness would script is learned inside the model. The framework predates this paper and does not discuss it.
What Is Not Established
The paper does not test large models, commercial tasks or real users. In the authors’ words: “Our experiments are limited to two tasks, small models, and one training run per condition, so additional runs may change some of our findings.”
The abstract says agents “exhibit emergent roles that complement each other”. The evidence the paper shows is agents claiming different branches of a search and settling on one broadcast each; it does not name fixed roles. The paper does not report a dollar cost for training or running a swarm.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
A University of Washington and Allen Institute for AI paper says swarms of small Qwen3 models can learn their own coordination, solving over 25 times more coordination problems than the strongest baseline at eight agents. The tests use two research tasks, and message deliveries rise from 12 to 240 as a swarm grows from four agents to sixteen, by our arithmetic.
A note on sourcing. Every result, figure and quotation is from the s3 paper by Jha and colleagues (arXiv 2611.08615, 5 November 2026), including its appendix tables; the X post is cited for its view count only. Message deliveries, team token allowances and the ratio between them are our arithmetic from the paper’s formula and settings. We have not reproduced the results.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
