# UW and Allen AI's s3 lets multi-agent teams learn their own organization

> Researchers from the University of Washington and the Allen Institute for AI introduce s3, a reinforcement learning method that lets agents with no preset roles learn to coordinate.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: Research
- Published: 2026-11-15T17:51:03+00:00
- Updated: 2026-11-15T18:10:24.732037+00:00
- Web page: https://ai-news.ikangai.com/story/uw-and-allen-ai-s-s3-lets-multi-agent-teams-learn-their-own-organization/
- Seen in: DAIR.AI (Elvis Saravia)
- Change: research (Research)

## What changed

The s3 method lets multi-agent teams learn their own coordination through reinforcement learning instead of relying on roles and orchestrators designed in advance.

Researchers from the University of Washington and the Allen Institute for AI have introduced s3, a multi-agent reinforcement learning method. Teams of agents learn how to divide work and share information, instead of following a structure their designers wrote in advance. The authors report that, with matched training and inference budgets, swarms evaluated with twice as many agents as in training solve over 25x more coordination problems than the strongest baseline. The method is described in an arXiv paper submitted on 5 November 2026.

The problem s3 targets is that multi-agent systems usually get their organization from harnesses designed in advance, such as an orchestrator that splits the task or a fixed set of roles. According to the paper's abstract, those static designs do not adapt to agent behaviors or task specifications. DAIR.AI's summary adds that such plans often duplicate work or flood agents with updates. In s3, a team is made of copies of one model, with no leader, no roles and no turn order. Each agent works on its own copy of the task. It decides for itself how long to work alone, when to message teammates and when to read their messages.

In training, all agents receive the same team reward, and every token an agent generates receives its team's advantage. Because task actions and communication are trained together, an agent can be rewarded for helping a teammate succeed. The communication behavior is learned, not scripted. The paper's abstract says the authors study two settings. The first is coordination, where agents must combine private information. The second is parallel exploration, where sharing discoveries can make search more efficient.

In the exploration setting, the abstract says communicating agents reduce repeated work by nearly 50% and outperform best-of-N by up to 9.4 percentage points. The 25x figure comes from a different experiment. The summaries do not give the absolute number of problems solved, so the size of that multiplier cannot be read against a base rate.

The experiments use small Qwen3 models, at 0.6B and 1.7B parameters. DAIR.AI's summary argues that the result shows coordination can be learned from a shared goal instead of hand-designed in a harness. The paper's claim is narrower. Agents learn different approaches, read each other's work, develop emergent roles that complement each other, and generalize to bigger teams.

## Why it matters

The result suggests coordination between agents can be learned from a shared goal rather than hand-designed in a harness, though the tests so far use very small models.

## Original source

- [(2611.08615) s$^3$: Simple Scaling of Swarms via Cooperative Multi-agent RL](https://arxiv.org/abs/2611.08615)
- [s$^3$: Simple Scaling of Swarms via Cooperative Multi-agent RL | DAIR.AI Academy | DAIR.AI Academy](https://academy.dair.ai/papers/s-3-simple-scaling-of-swarms-via-cooperative-multi-agent-rl-2611.08615)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "UW and Allen AI's s3 lets multi-agent teams learn their own organization", 2026-11-15, https://ai-news.ikangai.com/story/uw-and-allen-ai-s-s3-lets-multi-agent-teams-learn-their-own-organization/
