AI agents shine alone but fall out of sync with strangers: Oxford points to shared rules, not smarter models

By DiarioBitcoin · November 7, 2026.

A team at the University of Oxford has found that agents based on frontier language models consistently underperform when they must coordinate with partners they have never worked with. The paper is titled "Zero-Shot Coordination among LLM Agents." Its authors are Adrian Hayler, Shashank Reddy Chirra, Andrei Lupu, Johannes Forkel, Bidipta Sarkar and Siheng Feng. According to the source article, it was published on November 4, 2026.

The problem is called zero-shot coordination (ZSC): two agents must agree on how to act without having previously settled on a common strategy. To measure it, the researchers used minimalist environments based on a Dec-POMDP, a decentralized partially observable Markov decision process. In that framework, each agent decides with a limited view and must account for the fact that the other is also interpreting the situation on its own. One of the scenarios was a signaling game called Cat and Dog.

The study attributes the gap to two related weaknesses. The first is a limited understanding of how coordination works. The second is difficulty reasoning about what partners believe and why they make their decisions, that is, theory of mind (considering another's mental states).

The authors' proposal is coordination-friendly definitions (CFDs). These are principles written in natural language that are given to agents so they share conventions instead of relying on differing implicit assumptions. According to the results described, CFDs substantially improved coordination by favoring safe fallback actions and more consistent responses.

The team also tried generating them automatically through meta-learning, a technique that learns to produce useful strategies rather than receiving them pre-written. The CFDs discovered this way matched or outperformed handwritten ones.

There are clear limits. The environments are deliberately minimal, to isolate specific failure modes. The material provides no results on robotic warehouses, large-scale software projects or dynamically changing environments. According to the article, the study itself offers evidence under the conditions evaluated, not a guarantee of universal coordination.

**Our take.** The thesis we draw is simple: greater individual capability does not equal better teamwork. Until now, the conversation about agents has revolved around how well a single model solves a task. This work reminds us that the next bottleneck is social: interpreting the same instruction differently is not absurd from within each agent, but it produces incompatible actions.

The most valuable part is the kind of remedy. No new model or different architecture is needed: shared conventions and a safe option when the situation is ambiguous are enough. Broadly, this resembles what humans did with protocols: interoperability is rarely solved by individual intelligence, but by standards. It fits a thesis we have already argued: in the agent wars, the winner is whoever controls the "plumbing" (standards, integration, conventions), not just the smartest model.

There is also a governance reading. If agents from different teams and companies are going to interact, whoever deploys them needs to know what conventions they assume and what their default behavior is when in doubt. It ties into the gap we see between rapid adoption and minimal oversight. A "safe fallback action" is, at bottom, an auditable policy.

In the short term, this tempers dreams of heterogeneous agent swarms cooperating seamlessly; be wary of any demo that doesn't explain how the rules are agreed upon. In the long term, it is good news at heart: that the problem can be tackled with shared frameworks, even automatically generated ones, suggests that machine-to-machine coordination is a tractable engineering problem, not a wall. That is a necessary piece of the abundance we expect: systems that collaborate reliably with humans and with each other. That said, it remains to be shown that these rules hold up outside the lab.