Abstract for: AI Agents as Policymakers in Simulated Epidemics

AI agents are increasingly used for specialized tasks, yet their potential as computational models of policy decision-making remains underexplored. Testing whether generative AI agents can move beyond analytical support to actively make repetitive policy decisions in complex, feedback-rich environments is critical. We investigate this question in the context of epidemic policymaking, where decisions must balance public health outcomes with economic considerations under uncertainty. We develop a generative AI agent prompted as a city mayor, embedded within a simulated SEIR epidemic environment. Each week, the agent receives epidemiological data, retrieves recency-weighted memories, and sets business restriction levels. We examine performance across a 2×4 experimental design: two world models (policy-only vs. behavioral adaptation) and four agent configurations (base, knowledge intervention, ensemble, ensemble with knowledge), each repeated for 10 runs. The agent exhibits human-like reactive behavior, tightening restrictions as cases rise and relaxing them as risk declines. Providing brief systems-level knowledge about epidemic feedback loops reduces cumulative cases by approximately one-third. The ensemble with knowledge configuration achieves the strongest outcomes, cutting cumulative cases by half in World 1. Statistical analysis confirms that decisions are driven by recent case data and exhibit strong policy inertia. These findings illustrate how theory-informed prompting can shape emergent policy behavior in AI agents. These findings demonstrate that generative AI agents, when situated in structured environments and guided by minimal domain theory, can serve as powerful computational models for studying decision-making and policy design in complex social systems. Editing, proofreading