Oxford study finds LLM agents fail to coordinate with unfamiliar partners

Oxford researchers report that frontier LLM agents, including text and code agents built on GPT-5.2 and Gemini 3 Pro, fail to coordinate with partners they have not met before.

Across nine small cooperative games, agents paired with other models scored well below a guaranteed 5 points in the Cat & Dog game, each treating its own arbitrary convention as the obvious choice.

According to the report, a short shared statement of coordination principles, called a Coordination-friendly definition, led agents to reliably take the safe 5 points.