LLM Agents for Explainable Bridge Portfolio Maintenance Scheduling Under Uncertainty

Bridge maintenance scheduling is a complex decision-making problem due to stochastic deterioration, epistemic uncertainty in condition assessment, and financial constraints. Existing approaches rely on fixed rules or optimization frameworks that struggle to adapt to evolving information and provide limited explainability. To address these challenges, this work introduces a Large Language Model (LLM)-based agent for maintenance scheduling in bridge networks, embedded within a physics-informed simulation environment that models non-linear degradation and budget dynamics. The agent operates through structured prompts specifying system state, feasible actions, and operational rules, and outputs explainable decisions together with a machine-readable action plan. The experimental campaign evaluates: prompt configuration and memory mechanisms, five risk-aversion profiles, scalability with respect to bridge portfolio size, and long-horizon behavior. Across the tested scenarios (6–100 bridges, 7–60 years), the best-performing agent configuration achieved zero structural failures and retained final budgets up to 30% higher than the most conservative agent profiles. Prudent agents maintain consistently high structural reliability, whereas aggressive profiles exhibit reduced reliability and occasional failures, reflecting coherent behavioral differentiation. Relative to the tested heuristic and adaptive baselines, the proposed LLM agents achieve comparable performance in small and long-horizon settings, while avoiding the 20–60% failure rates observed for alternative strategies in large portfolios under tight decision horizons. The findings indicate that LLM agents hold promise as explainable and adaptable decision-making components for infrastructure life-cycle management. The results suggest that LLM-driven strategies may complement, and potentially enhance, existing scheduling and rule-based maintenance frameworks, motivating further investigation into real-world deployments and integration with domain-specific models.