Google's Earth AI agent exposes the steps behind its geographic answers
Expanded early access comes with benchmarks that distinguish useful tools and plans from proof that a final map or estimate is correct.
Google is expanding experimental access to an Earth AI agent that can plan and execute multistep geographic analysis, combining specialist models with maps, satellite data and an organization's own files. The more revealing part of its October 14 announcement is the evaluation: access to the right tools produces much of the improvement, while planning helps with less structured questions.
The research update describes early access through an expanded Ask Google Earth experience and an agent-to-agent API. These are research previews for a limited group, not a general release of an autonomous geographic analyst for every Google Earth user.
A map question becomes a sequence of operations
Consider selecting buildings for a community facility. A useful answer may require joining an address list with transport access, population data and a measure of rooftop solar potential. Each layer can have different coverage, dates and spatial resolution. A fluent paragraph cannot substitute for choosing and combining those datasets correctly.
Google's agent first prepares an analytical plan, then uses specialist capabilities in a separate workspace for large spatial data. It can accept inputs such as CSV files and geographic boundaries. Google says responses expose the plan and data sources so an expert can inspect the assumptions.
The separation of planning from execution matters because a geographic task often contains dependencies. A building must be located before its surroundings can be measured. A boundary must be defined before a population total has a clear meaning. If the wrong area is selected in the first step, accurate arithmetic later in the process can still produce an irrelevant answer.
What the comparison actually tests
The technical report compares four configurations: a standalone language model, the model with basic tools, a version with Earth AI tools but without the full planning guidance, and the complete agent. The researchers test both Gemini 3.1 Pro and 3.8 Flash.
This design is more useful than a single headline score because it asks where the improvement comes from. The report attributes the largest step to domain-specific skills and tools. On open-ended problems, planning and execution guidance provide a further gain for Flash. The conclusion concerns this system and these tasks; it does not establish that planning improves every agent by the same amount.
There is also an important distinction between checking an answer and checking a method. GeoKTools uses structured questions with reference solvers. GeoExpertQueries compares the agent's work with expert-authored workflows using an automated rater. A method can contain sensible steps while still producing an inaccurate final map or estimate.
Google explicitly treats the output as a starting point for expert review. That qualification belongs next to the results, because the product's usefulness depends on what the scores measure.
The released tests expose useful limits
The public benchmark repository includes 85 development queries for GeoKTools and 25 for GeoExpertQueries. Of the latter, 15 are described as beyond the current toolset. The held-out test sets remain private to reduce contamination.
Including questions the system cannot currently solve is a useful choice. An analyst needs to recognize when the necessary data or operation is unavailable, rather than manufacture a plausible result. A geographic answer can look convincing simply because it appears on a map; visual polish does not demonstrate that its underlying joins and measurements are valid.
The repository also distinguishes the report's scoring procedure from an additional deterministic offline scorer. Those are not automatically interchangeable ways to reproduce the same published number. Researchers comparing systems need to preserve the evaluation protocol as carefully as the questions.
This is a more concrete development than the broad funding ambitions discussed in our coverage of Google's AI science grants. Here there is an inspectable set of tasks, a description of the tools and a stated boundary around what the evaluation proves.
Earth AI's strongest proposition is reducing the work between a geographic question and a reviewable analysis. The obligation to check the area, dates, data coverage and final result remains with the people using that analysis. Exposing the intermediate work is valuable precisely because it gives them something more substantial to inspect than a confident answer.
