From Chat to Decision: a language-orchestrated and physics-grounded reasoning agent for infrastructure disaster resilience assessment and recovery decision support

Infrastructure resilience assessment and post-disaster recovery decision-making traditionally rely on fragmented engineering analysis models and expert-driven workflow configuration, limiting workflow continuity, transparency, and interactive adaptability in real-world engineering settings. To address these limitations, this study proposes a language-orchestrated and physics-grounded reasoning agent, termed Res-Agent, for disaster resilience assessment and recovery decision support. Res-Agent adopts a LangGraph-based state-driven orchestration mechanism, in which the large language model (LLM) serves only as a high-level controller for requirement understanding, task routing, workflow organization, and report articulation, while all numerical computation and physical reasoning are performed by encapsulated engineering tools. Hazard simulation, vulnerability assessment, functionality evaluation, resilience quantification, recovery optimization, and resource-constrained enhancement planning are integrated into a unified and traceable workflow. Structured memory maintains multi-round analytical consistency, while evidence cards and numerical consistency checks ensure that key quantitative outputs remain traceable to tool-derived results. A retrieval-augmented expert knowledge module constructed from academic publications, standards, and disaster reports provides interpretive support without interfering with numerical computation. To demonstrate the framework, UHVSR³-Agent is developed for a 750 kV ultra-high-voltage substation. Case studies show that the agent can interpret natural-language engineering requests, reconfigure analytical workflows, evaluate seismic functionality, generate resilience-oriented recovery and enhancement strategies, and maintain state consistency across interaction rounds. A prompt-based benchmark further confirms its reliability in tool routing, numerical traceability, and evidence-constrained reporting. This study provides a practical pathway for deploying LLM-based agents in high-consequence infrastructure resilience applications.