Multi-Agent Economic Simulations Using Large Language Models

Sammanfattning

Large Language Models have demonstrated remarkable capabilities in language understanding and generation, yet their potential for autonomous economic behavior in multi-agent systems remains largely unexplored. While traditional agent-based economic models rely on predetermined rules and simple heuristics, LLM-powered agents could theoretically exhibit emergent economic behaviors through natural language communication and reasoning. However, whether such agents can form functioning markets, discover prices through decentralized negotiation, and respond to environmental conditions in ways consistent with traditional economic theory is unknown. This research investigates whether LLM agents can sustain artificial economies through autonomous resource management and trade. Using a systematic experimental design, agents operated in a resource-constrained environment where survival depended on collecting resources and trading with other agents. The study manipulated resource availability, trading accessibility, and agent specialization patterns to test whether LLM agents exhibit three fundamental economic capabilities: forming stable markets with persistent trading activity, discovering prices endogenously through bilateral negotiation, and responding systematically to environmental constraints. The findings demonstrate that LLM agents can sustain basic economic functions over extended periods. Agents formed functioning markets with stable behavioral patterns, generating substantial trading activity across all conditions. They discovered prices through decentralized negotiation without programmed exchange rates, with accepted trades clustering around consistent exchange rate corridors (approximately 1.0–1.5 units of A per B), though whether this convergence reflects true economic equilibrium remains an open question. However, price convergence remained incomplete, with relatively low acceptance rates and persistent price variance indicating coordination challenges. Environmental conditions dominated outcomes, with resource availability explaining the majority of variance in economic success and survival. Resource scarcity prevented market formation entirely in extreme conditions, while abundance enabled survival but reduced price discovery quality. Trading accessibility influenced activity levels but not survival outcomes. Specialization showed context-dependent effects, providing advantages under resource abundance that diminished under scarcity. These results establish that current LLM agents can perform autonomous economic coordination through natural language but face significant limitations. They can form markets and discover prices, yet struggle with efficient coordination. Environmental constraints matter more than strategic sophistication in determining economic success. The findings provide baseline evidence for LLM agent economic capabilities while identifying coordination challenges and environmental sensitivities that must inform future multi-agent AI system design.