Large language models and AI agents in disaster-resilient infrastructure: concepts, applications, pathways, and challenges

This paper presents a systematic framework for integrating large language models (LLMs) and AI agents into disaster-resilient infrastructure research. By examining seven key task domains—knowledge anchoring, prediction, assessment, intelligent design and simulation, data augmentation, embodied intelligence, and decision support—we show how LLMs and AI agents can reshape traditional disaster-resilience workflows through enhanced multimodal perception, long-horizon reasoning, cross-stage coordination, and automated task execution. Two representative case studies further demonstrate the feasibility of this paradigm: (1) a multimodal building-damage assessment system capable of city-scale semantic reasoning and transparent diagnostic interpretation, and (2) a LangGraph-based reasoning agent for automated seismic resilience assessment and recovery decision-making of infrastructure systems. These cases highlight the potential of LLMs and AI agents to bridge fragmented data sources, orchestrate multi-stage analytical workflows, and support high-impact resilience applications. Despite these advances, key challenges remain, including data scarcity, reliability and consistency issues, deployment constraints, and ethical or security concerns. Addressing these limitations is essential for developing trustworthy, efficient, and scalable intelligent disaster-prevention systems. Overall, this work establishes a unified cognitive–action paradigm that offers forward-looking pathways for the next generation of intelligent and resilient infrastructure technologies.