A new AI agent doesn't just run a canned disaster model - it decides which scientific method actually fits the evidence in front of it.
Researchers built HazardWeaver, a system for automating hazard analysis - things like earthquakes, floods, and hurricanes - that has historically required human experts to pick the right scientific method for the data at hand. It uses three parts: a "Hazard Knowledge Compiler" that extracts the conditions under which a given method applies, a "Hazard Capability Graph" that maps which tools and models can actually run on the available data, and a "Weaver Agent" that selects a route, executes it, and reconsiders as new evidence comes in. The team tested it on a new benchmark of 141 scenarios spanning seven single-hazard types and four multi-hazard interaction classes, such as a quake triggering a landslide. HazardWeaver beat existing agent systems on the benchmark, with the biggest edge on cases where more than one scientific method could plausibly apply.
Most hazard-modeling software assumes a human already decided which method to run; this system tries to make that judgment call itself and show its reasoning for doing so. That matters because real disasters are often messy - an earthquake plus a landslide plus a flood - and the right analytical approach can shift as new readings arrive, which is exactly the situation that trips up existing automated tools, according to the study.
It's a university benchmark, not a FEMA deployment, and the code lives on GitHub, not in an operations center - the harder test is whether this holds up on messy, real-world data instead of curated scenarios.