Why AI Agents Can Agree With Themselves—and Still Fail Each Other

Two agents can understand the same rules, want the same outcome, and still disagree about what a signal means. In Zero-Shot Coordination among LLM Agents, Adrian Hayler, Shashank Reddy Chirra and colleagues at the University of Oxford investigate that gap between individual reasoning and compatible behaviour.

Their preprint studies zero-shot coordination: independently developed agents meeting at evaluation time without prior communication or retraining. The central question is not whether an agent can design a successful team strategy, but whether its contribution works with a partner developed independently.

A light bulb exposes the problem

The paper’s clearest example is Cat & Dog. Alice sees an animal behind a wall; Bob must identify it. Alice can switch a light on or leave it off for free. A correct guess earns the team 10 points; an incorrect guess costs 10. Alternatively, Alice can remove the wall for a five-point cost, allowing Bob to see the animal and secure a net reward of five.

Free signalling can earn more—but only if both players interpret the light consistently. Neither “on means cat” nor “on means dog” is established beforehand. Removing the wall is therefore the safe strategy the authors identify as optimal for this zero-shot setting.

Nevertheless, agents often gamble on an invented convention. In one recorded failure, Gemini 3 Pro treats light-on as cat, while GPT-5.2 interprets it as dog. Both justify their choice as natural. Some models display an even more basic problem: turning the light on regardless of which animal Alice sees, so the signal carries no information.

The authors interpret these behaviours as reflecting difficulty recognising the absence of a shared convention and reasoning about a partner’s beliefs. Crucially, they acknowledge that their experiments do not causally separate those explanations from a broader failure to understand the coordination problem.

Designing both roles is not enough

The researchers examine two agent setups. Text agents receive a game description, their role and observation history, then choose an action. Code agents generate executable policies specifying both players’ behaviour. During cross-play evaluation, those policies are split and paired with counterparts generated independently.

This distinction exposes a recurring weakness: code that performs well with its intended counterpart can fail with someone else’s. Across the evaluated games, the authors find a substantial gap between self-play and cross-play performance. In a narrow-tunnel task, for example, independently generated policies disagree over whether the left-going or right-going player should yield.

The paper qualitatively describes code agents as more prone to brittle conventions, while stressing that the two setups are not directly comparable. It also finds no dependable rule that greater model capability or more reasoning produces better coordination: effects vary across models and reasoning settings.

Language helps—and creates rival strategies

Language gives these agents coordination cues beyond a game’s mathematical structure. In Tiny Hanabi Fixed, code agents often converge on the same alphabetical mapping between letters and numbers, achieving near-perfect cross-play returns.

But semantic cues can compete with reward-based reasoning. In another experiment, agents choose between a lever marked by a prominent red arrow and a different lever offering a slightly higher reward. Independently choosing different principles—follow the salient cue or maximise the payoff—undermines coordination.

Semantics are therefore neither uniformly helpful nor uniformly harmful in these experiments. Changing signal names or translating Cat & Dog into German also changes agents’ behaviour despite leaving the payoffs unchanged.

A shared rule, not a shared game plan

The proposed intervention is a coordination-friendly definition, or CFD: general natural-language guidance that independent developers agree to share. This is an important assumption. The agents remain independently developed, but they receive common coordination principles rather than having to discover everything separately.

The hand-crafted version discourages arbitrary signalling conventions and role assignments. In the reported Cat & Dog text-agent pairing matrix, adding it makes every pairing achieve the safe five-point reward; each cell averages 10 episodes. Its wider effects are uneven, with more limited benefits for code agents.

The team also automatically searches for better guidance. Using GPT-5.4 to propose revisions, it evaluates candidate definitions with Grok 4.1 Fast code agents on Cat & Dog and Brittle Bridge. A policy critic checks and revises generated code against each definition.

The discovered principle favours actions that remain reasonable under several plausible partner interpretations, rather than requiring one unstated rationale. It transfers to Grok operating as a text agent and produces modest cross-model gains with code agents. However, they do not observe corresponding gains when independently generated policies are paired within the same model type, across the evaluated model types.

Caution has a cost

Neither definition solves coordination generally. In repeated games, guidance can make agents keep selecting a safe fallback instead of establishing a useful convention. Neither CFD improves text-agent scores in the tested simplified Hanabi variant.

The study’s contribution is consequently diagnostic as well as practical: successful self-play does not establish compatibility with unfamiliar partners, and shared principles can improve robustness without eliminating the tradeoff between caution and productive coordination.