Google Research has significantly advanced its Geospatial Reasoning agent, transforming it from an early prototype into a robust system capable of handling multi-step, planet-scale geospatial challenges. The latest updates include a new technical report detailing systematic evaluations across four benchmarks, including two newly developed benchmarks, GeoKTools and GeoExpertQueries. These benchmarks are now available for researchers to test their own geospatial reasoning agents.
The agent has been tested by real-world partners, including the World Health Organization (WHO) Regional Office for Africa, People in Need, GiveDirectly, the Milken Institute, and WPP. These organizations have used the agent for tasks such as disease outbreak investigations, disaster preparedness, and clean energy screening. Google is now expanding early access to the agent for both multi-step geospatial analysis and planetary prediction through Google Earth and Google Cloud.
The Geospatial Reasoning agent integrates four key capabilities: connection to Earth AI foundation models and planet-scale data, modular skills with upfront planning, the ability to incorporate user-provided datasets, and autonomous planetary prediction. The agent can now autonomously train custom predictive models, improving predictions of health outcomes by 12% over demographic baselines and enabling mapping at 8X finer spatial resolution.
Google's technical report compares four agent configurations, highlighting the importance of domain-specific tools and upfront planning. The evaluation shows that equipping the agent with Earth AI skills significantly improves performance across various benchmarks, except for UnivEARTH, where a standalone LLM already scores highly due to the availability of public web data.