From Rule-Based to Self-Improving Agents: The Evolution of AI Price Collusion
Sanya Shah
21/11/26
Traditionally, firms made pricing decisions manually. Over time, pricing systems evolved from rule-based algorithms to reinforcement learning and, more recently, large language model (LLM)-based agents that can read unstructured information, reason in natural language, and retain memory. A fourth generation — self-improving agents that could revise their own objectives — is anticipated but not yet in commercial use. This paper traces that evolution and examines how coordination works at each stage. Rule-based systems follow retaliation rules written by developers. Reinforcement learning agents learn to punish price cuts on their own, through trial and error. LLM agents can reach the same result: because rival firms’ agents often read the same public information and reason in similar ways, they raise prices together without exchanging a single message. The paper argues that each of these steps weakens the kind of evidence competition law depends on. Proof of collusion has moved from communication records, to patterns in pricing behaviour, and now to the decisions the algorithms themselves make. This evidence is hard to obtain, and hard to trust, since models do not always report their reasoning accurately. The paper then looks at how difficult it has become to assign legal responsibility and to detect coordination, and considers the tools regulators may need. These include detailed records of how pricing systems operate, the objective functions set for these algorithms, independent technical audits, rules set before deployment, explainable AI, and cooperation between competition authorities across borders.
