Beyond the prompt for pedestrian traffic and crowd behavior management: PedAgent, a Large Language Model empowered Agent with an integrated simulating-analyzing-optimizing functional chain

Pedestrian walking is one of the most fundamental modes of daily travel, yet recurrent pedestrian flow-related safety accidents, e.g., crowd crushes, highlight the urgent need for more intelligent management of high-density pedestrian facilities, for instance, urban subway stations. Existing pedestrian flow simulation models and empirically validated management policies provide valuable support, but they rarely satisfy the full-process requirements of crowd behavior management, i.e., simulation software can generate individual-level trajectories, but it usually lacks the ability to derive case-sensitive improvement policies from microscopic information; empirically validated policies are often tested under fixed boundary conditions, leaving their effectiveness uncertain in dynamic real-world environments. Recent advances in Large Language Models (LLMs) create opportunities to develop Agent s that support integrated simulation, analysis, and optimization for front-line users. However, high-density crowd dynamics involve multi-agent interactions among pedestrians, making a single prompt insufficient for guiding LLMs to address complex pedestrian flow problems and context-dependent user requests. To overcome these limitations, this study introduces PedAgent, a domain-specific Agent for pedestrian traffic management enabled by context engineering. PedAgent integrates LLM reasoning with a retrieval-augmented pedestrian knowledge base and a custom-developed Pedestrian Traffic Kit (PTK), establishing an end-to-end functional chain of simulating–analyzing–optimizing. Specifically, PedAgent can translate users’ natural language requests into executable simulation tasks and generates simulation-supported, qualitatively validated crowd management policies. Extensive case studies demonstrate PedAgent’s adaptability and generalizability in handling diverse user requests and managing multiple conflict hot-zones under different operational scenarios. Technically, PedAgent represents the first Agent for pedestrian traffic research and substantially improves the efficiency of simulation-assisted policy formulation by linking data generation, behavioral interpretation, and policy optimization. Practically, it bridges the gap between research-oriented theoretical findings and real-world policy formulation, thereby enhancing the intelligence of pedestrian traffic and crowd behavior management toward the development of safer and more efficient pedestrian facilities.