MAESTRO: Multi-Agent LLM System Turns Disaster Forecasts into Coordinated Action

Problem: From Forecast to Coordinated Action

Typhoon forecasts provide wind and rainfall predictions, but emergency responders need answers to operational questions: which areas may lose power, which roads will be cut, where to pre-position supplies. These questions span meteorology, hydrology, transportation, and emergency management. Even when each department has professional models, differences in data formats, terminology, and workflows keep information siloed. The challenge is turning prediction results into coordinated action faster.

Solution: MAESTRO Multi-Agent Architecture

A joint team from Zhejiang University, Huazhong University of Science and Technology, and the Zhejiang Provincial Department of Emergency Management proposed MAESTRO (published in npj Artificial Intelligence , DOI: 10.1038/s44387-026-00124-z). Instead of a single LLM answering all disaster questions, MAESTRO assigns three agent roles that mirror real institutional responsibilities:

Situational Awareness Agents monitor environmental changes and trigger tasks.

Execution Agents decompose requirements and orchestrate collaboration.

Domain Professional Agents invoke specialized models and knowledge bases (meteorological, hydrological, geological, etc.) to perform analyses.

Agents share state so downstream tasks inherit upstream results; independent tasks run in parallel, while dependent tasks follow a defined dependency graph. The system combines modular tool routing with a Reasoning–Action–Observation loop : retrieve knowledge, select models and parameters, check results, and adjust if needed. The LLM organizes computation rather than generating disaster conclusions directly from language.

Cross-checking across sources, conflict detection, and multi-round collaborative revision are built in. Data provenance and update logs are preserved, and final recommendations are submitted to the emergency command authority for approval and public release.

Evaluation: 100 Typhoon Scenarios

The team tested warning-level determination on 100 typhoon scenarios affecting Zhejiang: 10 historical events and 90 representative scenarios sampled from 5,000 WRF numerical simulations. They compared three configurations: MAESTRO (multi-agent + professional tools), a single-agent LLM baseline, and an expert team process.

MAESTRO : 98.4% average accuracy, 37.9 seconds average decision time.

Expert team : 250 seconds average decision time.

MAESTRO's runtime is ~15% of the expert process while achieving higher accuracy than the single-agent baseline.

The comparison focuses on warning-level judgment, not direct typhoon track prediction.

Beyond Warning Levels: Actionable Situation Reports and Task Plans

MAESTRO first lets each domain agent independently assess impacts (population, facilities, regions), then runs multi-round collaborative revision to resolve inconsistencies, producing a situation report with quantified results, risk maps, and source records.

In a blind evaluation by 20 provincial and municipal emergency professionals, the multi-agent collaborative reports matched or exceeded human-written baselines in clarity, structure, and data presentation, and scored highest on impact assessment and implementation feasibility.

For task planning, the team used 80 test tasks derived from real operational records . Task planning effectiveness and agent assignment accuracy both exceeded 90%; tool and parameter selection accuracy exceeded 83%.

In a cross-city resource scheduling case based on Typhoon Lekima, MAESTRO integrated wind/rain forecasts, flood and landslide risks, population exposure, and road conditions to generate a dispatch table, route map, and material list in ~120 seconds . The manual process took ~600 seconds. Output consistency with expert requirements was ~98%, and commanders judged only minor edits were needed.

Real-World Deployment: Zhejiang AI·1833R Platform

MAESTRO was integrated into the Zhejiang AI·1833R Disaster Assistant platform in mid-2024, connecting 8 industry departments, 11 municipal units, and 90 county-level emergency agencies.

During four typhoon responses (2024–2025), the system completed 80+ warning cycles .

Risk assessments verified by operators agreed with expert judgments ~95% of the time.

Action plans that previously required 9–12 experts and ~3 hours were generated in 3–5 minutes with only minor modifications needed.

Historical Replay: Typhoon Lekima 72-Hour Counterfactual

The team replayed the full 72-hour Lekima event. MAESTRO completed the "observation-to-warning" chain in 660 seconds versus an institutional baseline of ~5,940 seconds. The counterfactual estimate suggests that earlier, more spatially targeted warnings could have given ~180,000 people an extra 8 hours of evacuation lead time . This is a historical replay estimate, not an observed outcome.

Conclusions and Limitations

MAESTRO demonstrates that AI for Science should focus on organizing professional models, real-time data, and domain knowledge into collaborative, traceable workflows that bring scientific analysis closer to real decision-making needs. The system does not replace expert review; it still faces uncertainties from multi-source information, and cross-region deployment requires adaptation to local data, tools, and administrative processes. Future work must further evaluate actual disaster reduction effectiveness. For disaster AI, the key question is not only "how accurate is the prediction" but "can the analysis be turned into protective action more quickly under human supervision."

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
让科学分析更接近真实决策需求。