Google details Geospatial Reasoning agent and Earth AI benchmarks
Google Research reports evaluations of a geospatial agent combining Earth AI models with upfront planning and expands experimental access via Earth and an A2A API.
The story in focus
Google Research published a technical report on a Geospatial Reasoning agent that pairs Gemini with Earth AI foundation models and upfront planning. It reports benchmark gains from domain tools and planning, describes partner tests, and expands experimental access through Ask Google Earth and an agent-to-agent service.
- The agent combines Gemini with Earth AI models, Maps Platform, Earth Engine and Data Commons, plus user-supplied data.
- The report says Earth AI skills lift scores on GeoExpertQueries, GeoKTools and CBGB, while upfront planning matters for open-ended workflows.
- Google is releasing development splits of two new benchmarks and expanding limited early access to Ask Google Earth and an A2A API.
What the agent does
The experimental system connects Gemini to Earth AI foundation models covering weather, flood forecasting, population dynamics, remote sensing and satellite embeddings, alongside Google Maps Platform, Earth Engine, Data Commons and Dataset Search. Capabilities are packaged as modular skills the agent invokes only when a question requires them.
Before complex operations the agent builds a step-by-step plan, chooses fallback datasets where local coverage is sparse, and processes large map layers in a dedicated workspace. Users can also supply their own CSV, GeoJSON or cloud-table data for joining private locations with global models. The report says each response shows the plan and data sources so experts can check assumptions.
Evaluation and partner tests
The report compares four configurations across Gemini 3.8 Flash and Gemini 3.1 Pro on four benchmarks, including two new ones released in development splits: GeoExpertQueries, with expert-authored golden workflows, and GeoKTools, requiring correct use of one to four geospatial tools. It reports that domain-specific skills drive significant gains on GeoExpertQueries, GeoKTools and CBGB, while a standalone LLM still scored 93.6% on UnivEARTH, which the authors attribute to public availability of NASA source material.
The authors state that upfront planning is needed for open-ended GeoExpertQueries tasks to avoid unreliable assumptions or shortcuts. They note those scores come from an autorater comparing the agent to expert workflow steps rather than a single correct answer, and describe output as a strong first draft for expert review. Partners including WHO AFRO, People in Need, GiveDirectly, the Milken Institute and WPP tested the agent on outbreak mapping, cyclone preparedness, damage scoping and site screening.
Access and prediction
Google is expanding limited experimental access to a deep research update in Ask Google Earth and to a headless agent-to-agent API for developers, with recipes for connecting to Gemini. The agent also powers the recently announced Planetary Prediction Engine, which trains validated models from a user's local outcome data; the report cites a public health study where population and satellite signals improved outcome predictions by 12% over demographic baselines and enabled mapping at 8X finer resolution.
What did Google report about the value of planning in its geospatial agent?
The report says Earth AI skills alone sufficed for narrower benchmarks, but open-ended GeoExpertQueries tasks required upfront planning to keep multi-step analyses on track and prevent unreliable assumptions or web-search shortcuts.
Sources & further reading
Help keep the record accurate. Send a correction.
