Google moves geodata outside the context window and turns the map into an agent workspace

On October 14, Google published a technical report on its Geospatial Reasoning agent and expanded experimental access through Ask Google Earth and a headless A2A service. The agent first builds a plan, then selects models, datasets and spatial tools, processing large map layers in a separate workspace outside the context window.

The system connects Gemini with WeatherNext, Flood Forecasting, AlphaEarth Foundations, Earth Engine, Maps Platform and other sources. Users can add their own CSV, GeoJSON or cloud tables. Access remains limited and prospective users must submit an interest form.

Specialized tools outperform the language model alone

Google compared four configurations across four benchmarks. It started with a standalone model, then added ordinary web search and maps grounding, Earth AI skills and finally an upfront planning stage. The reported results show a substantial contribution from specialized tools on quantitative spatial tasks.

UnivEARTH was the exception, where the standalone model already scored 93.6%. Google notes that the underlying NASA articles and answers are available on the public web. That result may therefore measure retrieval of a known answer more than new spatial analysis.

A separate workspace solves a limit that prompting cannot

Planet-scale datasets do not fit sensibly inside a context window. The agent keeps the plan and conclusions in the conversation while processing map layers in a dedicated workspace. Each response shows the plan and data sources so an expert can change an assumption or substitute a dataset.

That architecture matters more than the chat interface. Geospatial work depends on projections, temporal resolution, local coverage and the choice of fallback data. The agent must know when to call a tool and also recognize when available data cannot support an answer.

Partner pilots are compelling case studies, not general proof

Google says WHO AFRO and the DRC's INRB used the system during the 2026 Ebola outbreak. The agent combined forest loss, mining locations and population data to identify 48 exposed settlements and more than 45,500 people at risk within minutes. People in Need analyzed 14 historical cyclone paths, while the Milken Institute screened more than 1,500 buildings.

These are cases selected and described by Google and its partners. GeoExpertQueries is also scored by an autorater against expert workflows rather than correct final answers. Google therefore describes the agent's output as a strong first draft for expert review.

Production value depends on audit, local data and correction time

The next test is not another demonstration on a global dataset. It is work with incomplete local records, conflicting sources and boundaries where a small error changes the number of people or buildings at risk. The useful measures will include an audit trail for each step and the time an expert spends correcting the output.

Limited early access must also show whether the A2A service preserves data provenance when connected to outside agents and private tables. If the source trail disappears between services, faster analysis will merely produce hard-to-trace errors sooner.

Lilith's verdict

Google has seated an agent above a map of the planet, but an expert must still draw the final red zone. In a flood or epidemic, a bad boundary is not cosmetic; it puts people on the wrong side of the map.