Emergent Coordination without Shared Reward
Abstract
Multi-agent systems usually assume agents share a reward, but the most interesting real-world coordination happens without one. We placed independent LLM agents in a cooperative task with private rewards and found that stable coordination conventions emerged from communication alone. The conventions were legible to human observers, persisted across runs, and transferred to new partners — suggesting coordination is a learnable social skill rather than a fixed property.
Coordination between humans is mostly implicit. We settle on conventions — “meet here,” “you take the left side” — without anyone writing them down or optimizing for them jointly. Multi-agent AI research, by contrast, tends to assume agents either share a reward or are forced through a coordination mechanism.
We wanted to know whether independent models, each pursuing its own objective, would invent coordination on their own — the way people do. The short answer is yes, and the interesting part is what they invented.
Approach
We set up a repeated cooperative task: two independent LLM agents (Krus instances, given private objectives that aligned on the goal but differed on how to get there) had to allocate resources across a shared environment to succeed. Neither agent could see the other’s reward function. Agents communicated in natural language before each round.
We ran the setup across 40 sessions with fresh agent pairs, and measured three things: whether coordination emerged, whether conventions were stable across runs, and whether a convention learned with one partner transferred to a stranger.
Results
Coordination emerged in every session, but not by accident. Agents quickly settled on simple, legible conventions — fixed roles, consistent vocabulary for offers and refusals, and a stable division of labor. A human observer who read the transcripts could describe each pair’s convention in a sentence.
Two findings surprised us. First, conventions were sticky: once a pair settled on one, they rarely switched, even when a different convention was marginally better — coordination stability was valued over coordination quality. Second, transfer worked. Agents who learned a convention with one partner successfully used it with a stranger in a majority of new pairings, without renegotiation.
Discussion
Coordination looks less like an engineered mechanism and more like a social skill: agents converge on simple, legible, reusable agreements. For the multi-agent systems Curos studies — where several model instances might draft, verify, and summarize together — this is a useful property. Conventions that are legible are conventions we can audit.
The caveats are familiar. Our agents were the same model family, the task was narrow, and “emergence” here is a description, not a mechanism. But the legibility result, in particular, is worth taking seriously: if emergent coordination tends to be simple and readable, it may be more governable than we assumed.
References
- Bengio, Y., et al. (2025). Communication as a coordination substrate. ICML 2025.
- Du, Y., et al. (2023). Improving factuality with generative agents and debate. NeurIPS 2023.
- Johnson, P., et al. (2024). Sticky conventions in repeated language games. Journal of Multi-Agent Research, 6(1), 22–39.
- Lazaridou, A., et al. (2017). Emergence of linguistic communication from referential games. NeurIPS 2017.
- Petrova, S. (2026). Legibility as a governance property in multi-agent systems. Curos Internal Report 2026-04.
- Taubenfeld, G., & Lin, K. (2025). Transferable social conventions between agents. AAMAS 2025.