Frontier AI agents struggle to coordinate even in simple tests, researchers find

Independently developed AI agents struggled to work together even in stripped-down tests, according to a new research paper. The authors trace much of the difficulty to understanding the coordination problem and reasoning about a partner’s beliefs—not simply navigating a complicated task. They also propose an approach intended to make coordination more reliable.

Zero-Shot Coordination among LLM Agents was submitted to arXiv on November 4, 2026, by Adrian Hayler, Shashank Reddy Chirra and five co-authors. It examines zero-shot coordination: situations in which independently developed agents must coordinate when put to the test.

Less complexity, a clearer coordination test

The researchers challenge two features of existing evaluations. First, work on language-model coordination often uses specialized scaffolding—the setup around the agents—which independently developed systems are unlikely to share in practice. Second, complex environments such as the game Hanabi make it difficult to identify precisely why coordination breaks down.

Their alternative pairs simple, general-purpose setups with minimal environments designed to isolate specific coordination challenges. The emphasis is diagnostic: rather than placing agents in a complicated setting and observing the overall result, the study tries to separate the underlying sources of failure.

Even under those controlled conditions, frontier language-model agents struggled. The authors identify limited understanding of the coordination problem and weak reasoning about partners’ beliefs as major causes. They also find that agents fail to use semantic information—the meaning available in language—effectively as an additional basis for coordination.

A proposed remedy, not a blanket solution

The paper proposes coordination-friendly definitions, or CFDs, as an approach to robust coordination. The authors report that these definitions can be discovered automatically, removing the need to engineer them manually. That is a separate contribution from diagnosing the failures: the work offers a possible intervention as well as an evaluation approach.

The abstract does not quantify the improvement from CFDs or name the frontier models evaluated. Its findings therefore describe a coordination weakness in the tested settings, not a ranking of products or proof that the proposed remedy works across every agent system.