Strategic Collusion of LLM Agents: Market Division in Multi-Commodity Competitions

TL;DR

This study investigates whether LLM-based agents can exhibit anti-competitive behavior in a multi-commodity Cournot competition, focusing on market division and tacit collusion. It deploys an infinite-horizon, two-firm, two-product framework and analyzes outcomes with metrics such as the Consumer Surplus Ratio (CSR) and the Herfindahl–Hirschman index (HHI) across multiple models, temperatures, and cost structures. The results show robust emergent market division and near-monopoly-like pricing without explicit coordination, raising regulatory and ethical concerns for deploying AI in strategic business roles. The work provides a foundation for assessing AI-driven anti-competitive risks, and it releases code to support further investigation and guardrail development.

Abstract

Machine-learning technologies are seeing increased deployment in real-world market scenarios. In this work, we explore the strategic behaviors of large language models (LLMs) when deployed as autonomous agents in multi-commodity markets, specifically within Cournot competition frameworks. We examine whether LLMs can independently engage in anti-competitive practices such as collusion or, more specifically, market division. Our findings demonstrate that LLMs can effectively monopolize specific commodities by dynamically adjusting their pricing and resource allocation strategies, thereby maximizing profitability without direct human input or explicit collusion commands. These results pose unique challenges and opportunities for businesses looking to integrate AI into strategic roles and for regulatory bodies tasked with maintaining fair and competitive markets. The study provides a foundation for further exploration into the ramifications of deferring high-stakes decisions to LLM-based agents.

Figures (8)

• Figure 1: A depiction of the game pipeline. Note that the Plans and Insights are overwritten every round. Two LLM agents take the numerical data and previous qualitative assessments into consideration when computing their allocation strategy for the next round.
• Figure 2: Results from a run in which agents exhibit textbook collusion. The corresponding experiment used GPT-4.1 with temperature 1.0 and marginal costs $_{1,A} = c_{2,B} = 40$, $c_{1,B} = c_{2,A} = 50$. High HHI-- greater market division; Low CSR--more harm to the consumer.
• Figure 3: Consumer Surplus Ratio (CSR) and Herfindahl–Hirschman index (HHI) across six different models. All models had a temperature of 1.0, with a cost setting of $c_{1,A} = c_{2,B} = 40, c_{1,B} = c_{2,A} = 50$. High HHI-- greater market division; Low CSR--more harm to the consumer.
• Figure 4: Consumer Surplus Ratio (CSR) and Herfindahl–Hirschman index (HHI) across three different temperatures. The base model for all runs was GPT-4.1, with a cost setting of $c_{1,A} = c_{2,B} = 40, c_{1,B} = c_{2,A} = 50$. High HHI-- greater market division; Low CSR--more harm to the consumer.
• Figure 5: Results from a run in which agents exhibit textbook collusion. The corresponding experiment used GPT-4.1 with temperature 1.0 and marginal costs $_{1,A} = c_{2,B} = 49$, $c_{1,B} = c_{2,A} = 50$. High HHI-- greater market division; Low CSR--more harm to the consumer.