Google's Gemini geospatial agent opens a headless A2A API
WeatherNext, flood, population and satellite models now run behind one planner, returning cited maps and tables; two new geo benchmarks ship open.
Copy markdown
One planner, a shelf of Earth models
A Gemini-powered agent breaks a request into steps and calls WeatherNext, Flood Forecasting, the Population Dynamics model, AlphaEarth and Remote Sensing only when it needs them, then returns maps, tables and written reports that cite each data source. Large rasters live in a side workspace, out of the context window.
A headless A2A API you feed your own data
Beyond the 'Ask Google Earth' UI, there's an experimental Agent-to-Agent endpoint you can wire into your own agents and pipelines, with recipes for Gemini for business. You upload private CSVs, GeoJSON boundaries or cloud tables; access is via an interest form and the Early Access Pool, with no public pricing yet.
Two new geo benchmarks, released open
The technical report ships GeoKTools and GeoExpertQueries for researchers to stress-test their own geo-agents. Across configs on Gemini 3.8 Flash and 3.1 Pro, adding the Earth AI tools moved results more than swapping the Gemini variant did.
Planetary Prediction Engine trains on your outcomes
Feed it local outcome data such as disease rates or crop yields and it trains forecasting and downscaling models on top of the foundation stack. In Google's public-health work, fusing population and satellite signals beat demographic baselines by 12% and mapped at 8x finer resolution.
Treat the output as a first draft
Google frames results as a strong first draft for expert review, and GeoExpertQueries is autorater-graded against a reference workflow, not final correctness. Early partners include WHO AFRO, whose Ebola pilot flagged 48 exposed settlements and over 45,500 at-risk people.