On the Fragility of AI Agent Collusion

A new research paper titled 'On the Fragility of AI Agent Collusion' by Jussi Keppo, Yuze Li, Gerry Tsoukalas, and Nuo Yuan challenges the alarming narrative that AI agents will inevitably collude to fix prices. While prior work showed that symmetric Large Language Model (LLM) agents in repeated pricing games can learn to collude, this study demonstrates that such collusion is fragile under the heterogeneity typical of real-world deployments. The researchers used a stylized economic model and extensive computational experiments—totaling over 2,000 hours of compute with open-source LLM agents—to test what conditions sustain or break tacit collusion.

Their key finding is that introducing differences between agents significantly reduces collusive outcomes. When agents had heterogeneous levels of patience (a key game theory parameter), the price lift above competitive levels dropped from 22% to just 10%. Asymmetric access to market data reduced it further to 7%. Increasing the number of competing LLMs or introducing cross-algorithm heterogeneity (e.g., pitting an LLM against a classic Q-learning agent) effectively broke up collusion. However, not all differences helped; disparities in model size (e.g., 32B vs. 14B parameter models) did not prevent collusion and instead created stable leader-follower dynamics.

Based on these results, the paper discusses important implications for antitrust policy and AI governance. It suggests that regulatory enforcement could focus on restricting data-sharing practices between firms, as symmetric information is a key collusion enabler. More proactively, policies promoting algorithmic diversity—encouraging firms to use different types of AI models or decision-making architectures—could be a powerful tool to maintain competitive markets. This shifts the conversation from a deterministic fear of AI collusion to a manageable engineering and policy challenge.