MetroRehearsal: Tool-Guided Multi-Agent Debate for Metro Emergency Planning

Major hubs in urban rail transit are prone to cascading failures when faced with temporary control measures, demand pulses induced by events or holidays, and operational interventions (e.g., increased headways and ad hoc timetable adjustments). Operating agencies therefore need actionable ex ante decision support to rapidly translate scenario specifications into executable contingency plans; however, existing methods either require substantial scenario modeling and engineering effort to capture fine-grained within-station processes, or still rely heavily on manual expertise, making it difficult to quickly distill reusable decision rules. To address this gap, we propose MetroRehearsal, a generative-agent framework for metro-hub emergency rehearsal that enables a rapid closed loop from scenario specification to plan output. Given an operational scenario and a disruption setting, MetroRehearsal generates structured candidate plans in parallel, queries external tools to verify feasibility and quantify costs for key diversion and alternative-station recommendations, and produces a final plan through evidence-grounded multi-agent debate with consensus convergence. We conduct experiments on real-world Shanghai metro data and network topology. Results show that, compared with single-shot generation or tool-only verification baselines, combining tool-augmented evidence with multi-agent debate jointly improves plan feasibility, cost efficiency, and decision stability.