EDSim: An Agentic Simulator for Emergency Department Operations

Abstract Emergency departments (EDs) face chronic crowding and complex patient flow challenges that traditional simulations struggle to capture. Conventional discrete-event or agent-based simulation can match high-level metrics, such as wait time distributions and throughput. However, they cannot reproduce the fine-grained behaviors, communications, and dynamic decision-making of real clinicians and patients. These micro-level interactions are often key to showing ED performance. We propose EDSim, an agentic ED flow simulator that uses large language model (LLM) agents to drive realistic, environment-aware interactions among artificial intelligence agents. EDSim offers a modularized patient journey from triage through treatment and discharge that supports customization, with virtual patients and healthcare providers who converse and make decisions in natural language based on dynamic conditions in the ED. The LLM agents are constrained by clinical rules and global EDstates, enabling both macro-level fidelity and micro-level insight. We have parameterized EDSim with aggregate historical data to reproduce daily arrival patterns and acuity mixes. Results show that EDSim’s baseline outputs align with historical wait time distributions stratified by triage acuity. Moreover, EDSim generates convincing individual conversations and behavior under potentially new workflows. This illustrates a new paradigm for healthcare operations research, combining data-driven agent-based modeling with LLM-generated behavior to provide a realistic, versatile testbed for improving ED operations. We demonstrate how analysts and hospital managers can use our tool to conduct what-if experiments in minutes, such as reallocating beds or staff to uncover bottlenecks and evaluate interventions. Our simulator can be used by both researchers and practitioners to explore ways to improve emergency care delivery.