MAESTRO: large language model-enabled cross-agency coordination for natural disaster early warning

Early warning systems are central to disaster resilience, yet impact-based alerts often falter under cross-agency fragmented mandates, semantic silos, and rigid approval chains that delay protective action. Despite advances in hazard forecasting, few platforms convert predictions into coordinated government response at scale. We present MAESTRO, a multi-agent system that mirrors institutional roles, mediates cross-agency semantics through natural-language reasoning and grounded tool-use, and integrates forecasts, impact models, and situational awareness under human oversight. Across 100 typhoon scenarios, MAESTRO matched expert alert levels in 98% of cases, reduced decision latency by over 85%, and produced reports rated clearer and more actionable by emergency professionals. In a 72-h replay of Typhoon Lekima (2019), earlier warnings enabled relocation of ~180,000 residents with eight additional hours of lead time. Deployed in a provincial government platform for multiple typhoon responses over 1 year, MAESTRO demonstrates one of the first AI-orchestrated early warning systems operating as live infrastructure—offering a scalable, inclusive pathway toward the global goal of Early Warning for All.