This paper presents AgentMonopoly, an experimental environment where LLM agents communicate, buy and sell properties, collect rent, and make turn‑based decisions. We evaluate four model families (each with two sizes) under three match‑up conditions: identical models, different sizes within the same family, and different families. Additionally, we assign one agent a persona—a textual description of the desired behavior—to see whether it can correct a tendency observed when the persona is absent. Results show that models differ markedly in trade frequency, rent waivers, cash and property management; larger models do not consistently win more games. Persona instructions alter the agents' wording and affect some actions, yet the changes do not uniformly align with the intended role across all decisions. These findings highlight that persona compliance should be judged by language, actions, and outcomes alike. The full codebase is open‑source, allowing the community to extend the comparative studies using the AgentMonopoly infrastructure.
Review