The Stanford Computational Antitrust Project announces the publication of “Smart Agent-Based Modelling with LLMs: Leveraging Large Language Models for a Better Understanding of Algorithmic Collusion” by Carlos Eduardo Veras Neves and Tanise Brandao Bussmann. The article appears in Issue VI of Stanford Computational Antitrust (pp. 1-31).

The paper introduces a Smart Agent-Based Modelling (SABM) framework within computational antitrust to simulate and detect the conditions that foster algorithmic collusion. The authors run Bertrand duopoly simulations in which LLM-driven agents stabilize prices above competitive levels without being explicitly instructed to do so. Simulations conducted in English and Portuguese show that linguistic context shapes outcomes. Communication between agents amplifies emergent behaviors, including the mimicking of concerns about collusion itself.

The contribution bridges market simulation and antitrust enforcement. By giving authorities an accessible tool to reproduce collusive conditions in silico, SABM offers a path to stress-test theories of harm before they materialize in real markets. The framework speaks directly to the open question of how regulators should address autonomous algorithmic collusion when no explicit agreement can be documented.

Thibault Schrepel, founder of the Stanford Computational Antitrust Project, stated:

“Carlos Eduardo Veras Neves and Tanise Brandao Bussmann contribute a practical instrument to a debate that has too often remained abstract. Their simulations confirm that LLM-driven pricing agents can drift into tacit collusion without being told to. Any agency thinking about how to monitor these markets should take note.”

The article is available for download over here.