Google ships a geospatial reasoning report and opens early access to Earth AI agents

Google Research published a technical report on October 14 and widened experimental access to its geospatial agent. The system, called a Geospatial Reasoning agent, has evolved from an early concept into a more capable experimental system built for multi-step geospatial reasoning across planet-scale data and the autonomous training of custom predictive models.

The design combines domain-specific foundation models with spatial tools and upfront planning. Those foundation models span remote sensing, weather, flood hydrology and population dynamics, and a Gemini-powered agent connects them so it can answer cross-modal questions instead of simply generating text.

The core evidence is a systematic evaluation across four benchmarks, including GeoKTools and GeoExpertQueries, two benchmarks the team built and is releasing so other researchers can test their own geospatial reasoning agents. Google says the results quantify the value of Earth AI's specialized tools and planning over out-of-the-box LLMs.

For real-world testing, Google invited a small group of social impact and commercial partners over the past few months, including the WHO Regional Office for Africa, People in Need, GiveDirectly, the Milken Institute and WPP, and put the agent to work on workflows such as disease outbreak investigations, disaster preparedness and clean energy screening.

Access is expanding on two fronts: an experimental deep research update for Ask Google Earth, and an early-access agent-to-agent, or A2A, API with integration recipes for developers building custom workflows.

Google released the benchmarks, a microsite and an interest form alongside the report. The new benchmarks give researchers a comparable yardstick, while the A2A route lets organizations embed geospatial analysis into their own systems rather than working only through Google's interface.

Why it matters: planetary analysis has long depended on experts and bespoke pipelines, so packaging foundation models, tool use and multi-step planning into an agent is a test of whether model capability can be turned into deliverable workflows.

What to watch next is how the early-access partners perform. If disaster response, public health and energy screening tasks run reliably, and outside teams reproduce Google's benchmark claims, the case for specialized geospatial agents over general models gets much stronger.