AI Pricing, Agent Heterogeneity, and Collusion

Recent studies find that firms using similarly capable LLM pricing agents can reproduce classical algorithmic collusion results. In practice, however, firms’ agents differ in multiple dimensions, including algorithm type (e.g., LLMs vs. reinforcement learning), model sophistication, size, and data access. We provide one of the first joint theoretical and experimental analyses of how such heterogeneity affects algorithmic collusion. We develop a model showing that differences in sophistication and data access weaken collusion sustainability. We then test and validate these predictions in a lab framework using open-source LLM agents. Leveraging this framework, we further show that both a greater number of agents and greater algorithmic diversity reduce collusion, while differences in model size alone do not. These findings highlight agent heterogeneity as a new factor for firms and regulators managing collusion in the age of agentic AI.