A University of Oxford paper tested AI agents from five model makers in a simple two-player game where a safe move guarantees 5 points (arXiv 2611.06409, posted November 4). Agents built on different models averaged -0.51 points together, against 9.75 when each model was paired with its own policies.
The authors write that “frontier LLM agents struggle to coordinate, largely due to limited understanding of the coordination problem and weak reasoning about their partner’s beliefs”. A short written rule against arbitrary conventions lifted the mixed-model average to 4.46.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how a harness around models shapes what they deliver. It teaches the general idea only and does not discuss this study.
The key insight: An Oxford paper posted on November 4 found that code agents built on five frontier models averaged 9.75 points in a simple coordination game when paired with their own model’s policies and -0.51 with other models’, below the 5 points a safe move guarantees.
The paper, Zero-Shot Coordination among LLM Agents, is by Adrian Hayler and Shashank Reddy Chirra, who contributed equally, with Andrei Lupu, Johannes Forkel, Bidipta Sarkar, Siheng Feng and Jakob Foerster of BOLD at the University of Oxford.
Zero-shot coordination means independently built agents must work together at test time with no agreed rules. The researchers ran nine small cooperative games and two kinds of agents: text agents, which pick moves turn by turn, and code agents, which write a full strategy as code.
In the Cat & Dog game, one player sees an animal and can end the game for 1 point, pay 5 points to remove a wall so the partner can see the animal, which guarantees a joint 5, or switch on a light for free. The partner then guesses: a correct guess pays 10 and a wrong one costs 10. The light only helps if both players read it the same way, and there are two equally valid ways to read it.
Mean Cat & Dog score of code agents paired with policies from the other four models, without (grey) and with the paper’s short coordination-friendly rule. The safe move guarantees 5. Data: Hayler, Chirra et al., arXiv 2611.06409, Table 6. Changes are our arithmetic.
What Was Posted on X
New Oxford paper finds that frontier LLM agents fail to coordinate with unfamiliar partners because each one treats its own made-up rules (arbitrary convention) as the obvious choice.
Agents from different companies increasingly must work together with no agreed rules. Past studies used shared scaffolding or complex games that hide why coordination fails.
They tested text and code agents built on models like GPT-5.2 and Gemini 3 Pro in 9 small cooperative games. In Cat & Dog, a safe move guarantees 5 points, while a signal convention pays 10 only if both agents pick the same one.
Paired with other models, text agents scored well below that safe 5, each calling its own arbitrary convention natural. A short shared statement of coordination principles, called a Coordination-friendly definition, led agents to reliably take the safe 5.
If your agent must work with agents you did not build, tell it to avoid unstated conventions and prefer actions any partner reads correctly.
A post by Rohan Paul at 18:45 UTC on 7 November summarising the paper: it says agents built on models such as GPT-5.2 and Gemini 3 Pro were tested in nine small cooperative games, and that text agents paired with other models scored well below the safe 5 in Cat & Dog. It describes the coordination-friendly definition as a short shared statement of principles. The closing advice to avoid unstated conventions is the poster’s view.
The Same-Model vs Mixed-Model Gap
For code agents in Cat & Dog, the paper reports self-play scores, where a model is paired with its own independently generated policies, and cross-model scores, where it is paired with policies from the other four models. The models were GPT-5.2, Grok 4.1 Fast, Claude Opus 4.5, Gemini 3 Pro and DeepSeek V3.2 Terminus.
Self-play averaged 9.75. Cross-model play averaged -0.51, which is below the 5 points the safe move guarantees. By our arithmetic, that is a gap of 10.26 points. GPT-5.2 scored -3.03 with other models’ policies and Gemini 3 Pro -1.63, while Claude Opus 4.5 scored 1.20, according to the paper’s Table 6.
The paper says Gemini 3 Pro consistently read the light as meaning cat, which let it coordinate with itself but not with others. In one failed episode the paper shows, Gemini 3 Pro and GPT-5.2 “independently arrive at incompatible strategies”.
Code agents also leaned on alphabetical ordering, the paper says. Some mapped cat to light on because cat comes before dog, while others treated off and on as 0 and 1. The authors note the ordering can invert in another language: “Hund (dog) precedes Katze (cat) in German”.
More Reasoning Did Not Fix It
Text agents struggled too. In the paper’s Cat & Dog cross-play tests, GPT-5 mini averaged 3.23 points at low reasoning effort and 0.86 at high effort, and Grok 4.1 Fast averaged -0.75 at medium effort, according to the paper’s Table 3.
The paper also reports how often frontier models, playing the first role, took the safe move of removing the wall. At high reasoning effort, Claude Opus 5 did so 74% of the time and Grok 4.5 49%, while GPT-5.6 Sol used the light as a signal 89% of the time.
A Short Rule Helped
The authors propose what they call coordination-friendly definitions: short instructions telling an agent to avoid moves that rely on arbitrary conventions or role assignments. Their hand-written version asks whether a signalling convention “could be replaced by a different, equivalent convention”.
With that rule, the code agents’ cross-model Cat & Dog average rose from -0.51 to 4.46, and GPT-5 mini text agents at high effort rose from 0.86 to 4.92. The paper says the rule had a weaker effect for code agents in most other games, and in the ad-hoc convention games it could be “too restrictive”.
The researchers also used an automated search to discover a shorter rule, which reads: “Prefer actions that would still be reasonable under several plausible partner interpretations of the shared observation”. The paper says it is roughly two-thirds shorter than the hand-written version.
The Business Lever
The paper cites coding tools such as Cursor, where agents from different model providers work on the same code, as a setting that needs this kind of coordination. When agents from different vendors interact without agreed rules, this study finds that each tends to treat its own guess as the obvious one.
As we read the paper, the failure it isolates is not a lack of capability: each model could coordinate with its own policies, and the trouble came from each treating its own arbitrary choice as the shared one.
As we read the results, a few sentences of instruction changed outcomes more than higher reasoning effort did for GPT-5 mini, whose cross-play score fell from 3.23 at low effort to 0.86 at high effort and rose to 4.92 with the rule.
As we read it, the games are deliberately small and the authors call strong results a necessary but not sufficient condition, so the study shows a failure mode rather than how deployed products behave together.
Hayler, Chirra et al., arXiv 2611.06409
“Our results show that even in these controlled settings, frontier LLM agents struggle to coordinate, largely due to limited understanding of the coordination problem and weak reasoning about their partner’s beliefs.”
Three Implications
FOR TEAMS MIXING MODEL VENDORS The paper cites coding tools such as Cursor, where agents powered by different LLM providers are used together, as a setting that needs this kind of coordination.
FOR AGENT BUILDERS The paper reports that its hand-written coordination-friendly rule raised the code agents’ cross-model Cat & Dog average from -0.51 to 4.46, and that an automatically discovered rule was roughly two-thirds shorter.
FOR RESEARCHERS The paper says the rule had a weaker effect for code agents beyond Cat & Dog and the Semantic Selection Game, and could be too restrictive in the ad-hoc convention games.
The essay describes “an operating system that wraps your AI”, “a harness that compresses your thinking, your team’s feedback, and the model’s raw capability into something replicable.”
As we read it, the paper’s coordination-friendly rule is a small piece of that harness: a few sentences added around the same models changed whether agents from different makers could work together. The essay does not discuss this study.
What Is Not Established
The games are small and designed to isolate one failure at a time. The authors write that strong results in them are “a necessary but not sufficient condition for effective real-world coordination”. The study does not show how deployed products from these companies behave together, and this piece makes no forecast.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
In an Oxford study of AI agents from five model makers, code agents scored 9.75 on average with their own model’s policies in a simple coordination game and -0.51 with other models’, below the 5 points a safe move guarantees. A short written rule against arbitrary conventions raised the mixed-model average to 4.46.
A note on sourcing. Figures and quotations are from the paper by Hayler, Chirra et al. posted to arXiv on 4 November 2026 (2611.06409). We did not rerun the experiments. The gap and point changes are our arithmetic from the paper’s Table 6.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
