AI Agents as Policymakers in Simulated Epidemics

TL;DR

The paper investigates whether generative AI agents can serve as repetitive policy decision-makers in epidemic settings. It embeds a memory-augmented LLM policymaker in a SEIR-based simulation, comparing base, knowledge, ensemble, and ensemble-with-knowledge configurations across two world models. The key finding is that providing systems-level knowledge about epidemic feedback loops substantially improves decision quality and stability, with ensemble-with-knowledge delivering the strongest outcomes in the simpler world; behavioral adaptation attenuates gains but knowledge remains beneficial. The work demonstrates that theory-informed prompting and memory scaffolds can guide emergent policy behavior in AI agents and offers a controlled framework for studying decision-making and policy design in complex social-epidemic systems.

Abstract

AI agents are increasingly deployed as quasi-autonomous systems for specialized tasks, yet their potential as computational models of decision-making remains underexplored. We develop a generative AI agent to study repetitive policy decisions during an epidemic, embedding the agent, prompted to act as a city mayor, within a simulated SEIR environment. Each week, the agent receives updated epidemiological information, evaluates the evolving situation, and sets business restriction levels. The agent is equipped with a dynamic memory that weights past events by recency and is evaluated in both single- and ensemble-agent settings across environments of varying complexity. Across scenarios, the agent exhibits human-like reactive behavior, tightening restrictions in response to rising cases and relaxing them as risk declines. Crucially, providing the agent with brief systems-level knowledge of epidemic dynamics, highlighting feedbacks between disease spread and behavioral responses, substantially improves decision quality and stability. The results 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.

Figures (5)

• Figure 1: Conceptual architecture of an AI agent as a policymaker (left) coupled with a mechanistic model of simulated epidemic (right). Note: Every week the policymaking agent observes recent cases and retrieved memories, outputs a business-restriction level; the SEIR environment advances affected by the policy, internal dynamics of the spread of virus, and in one of the world-scenarios with public behavioral response (dotted lines).
• Figure 2: World 1 (policy feedback): Mean trajectories over 10 simulation runs. Top panel: reported cases; bottom panel: reduction in transmission implied by policy (lower = stronger suppression).
• Figure 3: World 1 (policy feedback only): Performance across agent configurations. (a) Cumulative cases indexed to single-agent baseline; (b) mean reduction in transmission (lower = stronger suppression); (c) cumulative prediction error indexed to baseline.
• Figure 4: World 2 (policy + behavioral adaptation): Mean trajectories over 10 simulation runs. Top panel: reported cases; bottom panel: reduction in transmission implied by policy (lower = stronger suppression).
• Figure 5: World 2 (policy + behavioral adaptation): Performance across agent configurations. (a) Cumulative cases indexed to single-agent baseline; (b) mean reduction in transmission (lower = stronger suppression); (c) cumulative prediction error indexed to baseline.