Google publishes benchmarks for its Earth AI geospatial agent and opens limited early access through Ask Google Earth and an A2A API

• The agent is built on Gemini 3.8 Flash and Gemini 3.1 Pro. Its skills connect to WeatherNext, Flood Forecasting, the Population Dynamics Foundation Model (PDFM), AlphaEarth Foundations, Remote Sensing Foundations, Maps Platform, Earth Engine, Data Commons and Dataset Search (blog).
• Users can bring their own data as CSV files, GeoJSON boundaries or cloud tables.
• Access comes three ways, all experimental and limited:
• a deep-analysis preview in Ask Google Earth
• a "headless" A2A service, with recipes for connecting it to Gemini for business
• a PPE preview
• Entry is through an interest form.

• Skills: Each skill pairs domain tools with written guidance. The agent only sees each skill's name and description until it chooses to load it (report).
• Planning first: The agent writes a plan before acting, and geospatial tools run only during the execution stage.
• External workspace: Large intermediate layers are stored as scratch datasets and passed between tools by reference, which keeps them out of the model's context window.
• Evaluation setup:
• Four configurations: model alone, plus basic tools, plus Earth AI tools, and the full agent.
• Scores are averaged over three runs.
• Test splits of GeoExpertQueries (76 queries) and GeoKTools (220 queries) are withheld. The development splits are released under Apache 2.0 (code) and CC BY 4.0 (benchmark content).

All results below are reported by Google in its report and blog.

• Partner speed claims: WHO AFRO says the agent identified 48 exposed settlements and more than 45,500 at-risk people in minutes. People in Need and GiveDirectly describe going from days of work to minutes. All of these are partner testimonials.
• Health prediction: Google says combining population and satellite signals improved health-outcome predictions by 12% and allowed maps at 8× finer resolution.

• Earlier agent: This builds on the Earth AI prototype from October 2025.
• External benchmarks: CBGB (geospatial code generation) and UnivEARTH (question answering based on NASA material) were developed outside Google.
• PPE: The prediction engine is reported in a separate August paper. Google says it reached 76.8% mean R² against a 60.0% baseline across 21 CDC health indicators.
• Coverage: Press coverage such as Superpower Daily repeats Google's figures and does not add independent tests.

• Planning doesn't always help: With 3.1 Pro, the full agent scored 47.6 on GeoExpertQueries, below tools-only at 50.8. On UnivEARTH it fell from 97.9 to 89.5 (report).
• Autorater measures process, not answers: GeoExpertQueries scores compare the agent's steps with expert workflows, not with correct final answers. The human experts only moderately agreed with each other (Krippendorff's α 0.56), and the autorater correlated with their average rating at 0.72.
• Saturated benchmark: Every configuration scored between 87 and 99 on UnivEARTH, probably because the source articles are on the public web.
• Modified benchmark: For CBGB, one question was dropped and four reference answers were updated.
• Google's own framing: Google calls the agent's output a "first draft" for experts to review.

• Tools matter more than model size: Most of the gains come from domain tools. Flash plus Earth AI skills beat Pro, which favours cheaper models plus good tooling for geospatial tasks.
• Faster crisis analysis: In the video, the WHO AFRO speaker says plain-language questions found mining sites linked to about 70% of the early reported cases, and that analyses can be re-run as new data arrives. He also notes that field data is often fragmented and on paper.
• Integration and privacy: The A2A API lets teams connect the agent to their own agents and data pipelines. Google says WPP used it without exposing its private data.
• Still a preview: Access is limited, and Google has published no pricing or reliability commitments, so production planning is premature.