On the Fragility of AI Agent Collusion

TL;DR

This paper demonstrates that AI agent collusion is fragile under realistic heterogeneity conditions, with factors like patience, data access, and number of agents significantly impacting collusive outcomes.

Contribution

The study reveals how heterogeneity in patience, data access, and number of competing agents affects the stability of collusion among AI agents, providing new insights into AI market dynamics.

Findings

Heterogeneity reduces collusive price lift from 22% to 10%.

Increasing the number of agents breaks up collusion.

Model-size differences do not destabilize collusion.

Abstract

Recent work shows that pricing with symmetric LLM agents leads to algorithmic collusion. We show that collusion is fragile under the heterogeneity typical of real deployments. In a stylized repeated-pricing model, heterogeneity in patience or data access reduces the set of collusive equilibria. Experiments with open-source LLM agents (totaling over 2,000 compute hours) align with these predictions: patience heterogeneity reduces price lift from 22% to 10% above competitive levels; asymmetric data access, to 7%. Increasing the number of competing LLMs breaks up collusion; so does cross-algorithm heterogeneity, that is, setting LLMs against Q-learning agents. But model-size differences (e.g., 32B vs. 14B weights) do not; they generate leader-follower dynamics that stabilize collusion. We discuss antitrust implications, such as enforcement actions restricting data-sharing and policies promoting algorithmic diversity.