Google Research details Earth AI agent for planet-scale analysis

What Google published

On 14 October 2026, Google Research described how a Geospatial Reasoning agent, built on its Earth AI models and Gemini, can handle complex spatial analysis that standalone language models struggle with. The blog post by Shravya Shetty and Luke Barrington accompanies a new technical report ^[1].

The agent connects Gemini to Earth AI foundation models, including WeatherNext, Flood Forecasting, the Population Dynamics Foundation Model, AlphaEarth Foundations and Remote Sensing Foundations. It also draws on Google Maps Platform, Google Earth Engine, Data Commons and Google Dataset Search ^[1].

How it works

Earth AI capabilities are packaged as modular skills that the agent calls only when a question needs them. Before running complex operations, the agent builds a step-by-step plan and processes large map layers in a dedicated workspace outside the model's context window. Every response shows the plan and the data sources used so experts can check assumptions ^[1]. Users can also bring their own data as CSV files, GeoJSON boundaries or cloud tables ^[1].

The agent also powers a Planetary Prediction Engine, which trains machine learning models on a user's local outcome data. In Google's public health research, combining population and satellite signals improved predictions of health outcomes by 12% over demographic baselines and enabled mapping at 8X finer spatial resolution ^[1].

The evaluation

The report compares four set-ups across Gemini 3.8 Flash and Gemini 3.1 Pro, from a standalone model to the full agent with upfront planning. It uses four benchmarks, two of them new: GeoExpertQueries and GeoKTools. Google is releasing the development splits of both ^[1].

Google's two takeaways are that domain-specific skills are essential for spatial reasoning, and that upfront planning helps with open-ended, multi-step work. One exception: on UnivEARTH, even a standalone model scores 93.6%, likely because the source articles and answers are on the public web ^[1].

Google acknowledges limits. GeoExpertQueries scores come from an autorater comparing the agent's steps with expert workflows, so Google treats the output as a strong first draft for experts rather than a final answer ^[1].

Partner testing

Partners tested the agent on real tasks. WHO's Regional Office for Africa said it used the agent at the start of the 2026 Ebola outbreak in the Democratic Republic of the Congo to map remote mining corridors, identifying 48 exposed settlements and more than 45,500 at-risk people in minutes, a process Google says would normally take weeks ^[1]. People in Need used it to analyse 14 historical cyclone paths, and GiveDirectly used it to scope crises such as Tropical Cyclone Maila. The Milken Institute screened more than 1,500 candidate buildings for community resilience hubs, and WPP used it for geospatial marketing ^[1]. These accounts come from Google and its partners.

Access and what it means

Google is offering early access through an experimental deep-research update to Ask Google Earth and an experimental agent-to-agent API for developers. Interested users can apply through an interest form ^[1].

For businesses, the message is that location-heavy questions, such as site selection, risk exposure and logistics, may soon be answerable in minutes rather than days. The caveat is that these are experimental tools, results need expert review, and the evidence so far is Google's own.

Source: Google Research — Putting Earth AI to work: Geospatial agents for our planet's biggest challenges