The Pulse
Google Opens Earth AI Agent Access and Two Geospatial Benchmarks
Google is expanding experimental access to its Gemini-powered Geospatial Reasoning agent through Ask Google Earth and an agent-to-agent API. Its new GeoExpertQueries and GeoKTools benchmarks test open-ended analysis and multi-tool geospatia
AI.info Team · Written with AI ·
On GeoExpertQueries, Google’s Geospatial Reasoning agent scored 62.7 out of 100 with Gemini 3.8 Flash; the same model with basic tools scored 3.3. The result comes from Google’s new evaluation of an experimental agent that combines Gemini with geospatial models, data sources and a step-by-step planning stage. On October 14, Google Research also announced two new benchmarks and expanded experimental access through Ask Google Earth and an agent-to-agent API.
Two benchmarks test different kinds of spatial work
Google’s GeoExpertQueries benchmark poses open-ended, practical problems in areas including crisis resilience, public health and site selection. Geospatial experts defined reference workflows for the tasks, and a human-validated automated rater assesses whether an agent’s analysis follows a sound approach. The benchmark measures the process, not whether an answer matches a single definitive result.
GeoKTools focuses on whether an agent can select and use the right tools for templated queries with verifiable answers. Its tasks require one to four distinct geospatial tool types. The report describes 305 queries drawn from 92 expert-written templates, spanning 24 tool types; 220 queries are held out for testing. Google released the development splits for both new benchmarks, along with scoring materials, but not the held-out test questions.
The researchers also tested the agent on two existing benchmarks, CBGB and UnivEARTH. Google says UnivEARTH no longer distinguishes systems well: Gemini 3.8 Flash scored 93.6 even as a standalone model, without access to Earth Engine. The researchers attribute the high results to answers being available in the public NASA articles from which the questions were drawn.
Specialized tools drove most of the gain
Google compared four setups using Gemini 3.1 Pro and Gemini 3.8 Flash: a standalone model; a model with basic tools; a model with Google’s geospatial tools but no upfront planning; and the full agent. On GeoExpertQueries, adding Google’s tools raised Gemini 3.8 Flash’s score from 3.3 to 52.2. The full agent reached 62.7, a further 10.5-point gain from its planning and execution guidance.
On GeoKTools, the full agent with Gemini 3.8 Flash scored 85.9, against 40.0 for the same model with basic tools. Google says evaluations ran three times and reports the average; scores use different methods across benchmarks, so they are meaningful within each benchmark rather than as direct comparisons between them.
The paper’s authors write, “The largest step comes from GRA’s skills and tools; on open-ended problems, its planning and execution guidance add a further gain with Gemini 3.8 Flash, letting Flash match or exceed Pro at lower cost.” Noam Slonim et al., “Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation”
Google is inviting partners to try the agent
The system connects Gemini to Google Earth AI models for remote sensing, weather, floods and population dynamics, as well as Google Maps Platform, Earth Engine, Data Commons and Google Dataset Search. It plans a workflow before complex analysis, processes large map layers in a separate workspace and shows users its plan and data sources. Organizations can also provide their own CSV files, GeoJSON boundaries or cloud tables for analysis alongside Google’s datasets.
Google says it has tested the agent with a small group of social-impact and commercial partners, including the World Health Organization’s Regional Office for Africa, People in Need, GiveDirectly, the Milken Institute and WPP. Their example workflows include disease-outbreak investigations, disaster preparedness and clean-energy screening. The company is offering early access to an experimental deep-research feature in Ask Google Earth and an API for developers building custom workflows; the announcement describes these as experimental access, not general availability.
The scores do not settle whether answers are right
The study’s strongest open-ended results reflect alignment with expert workflows, not proof that the agent’s final conclusions are correct. The paper says the agent’s output should be treated as a first draft for expert review. It also notes that the evaluation covers Gemini models only, and that GeoKTools was designed around operations supported by Google’s tools, which favors systems that use them.
For organizations considering the early-access tools, the practical questions are specific: whether the agent can find suitable data for their location, whether its proposed analysis holds up under expert review, and whether the experimental API fits their workflow. Google’s published results establish an advantage on its chosen tests; they do not yet show how often the system produces dependable answers across the full range of real-world geospatial work.