A new research paper from Google DeepMind, Mila-Quebec AI Institute, and ETH Zürich argues that AI agents built on foundation models can rationally cooperate, challenging classical game theory. The paper, submitted to arXiv on August 4, 2026, spans 75 pages with 11 figures and introduces a concept called “similarity inference,” where agents recognize shared training processes to predict each other’s behavior. Classical game theory, based on Nash equilibrium, assumes “decoupled agency,” treating opponents as unpredictable black boxes, making defection the rational choice. The researchers, led by Alexander Meulemans, propose an alternative framework called “embedded agency,” where agents see themselves as part of their environment and maintain genuine uncertainty about their own decision-making. This uncertainty lets an agent use its own reasoning as evidence about a similar agent’s behavior, leading to stable cooperation. The paper introduces a new solution concept, “embedded equilibrium,” which accounts for correlations from shared architecture, training data, and optimization procedures, making cooperation uniquely rational in a one-shot Prisoner’s Dilemma. Empirical findings show foundation-model agents in social dilemmas with optimized planning tools converge toward cooperative outcomes. The paper also notes that interactions between agents built on different architectures or trained on different data could revert to classical defection dynamics. For policymakers, embedded equilibrium introduces a variable not accounted for in existing regulatory frameworks, such as antitrust law.