Google Releases Earth AI Agent Framework with Agent-to-Agent Service and Spatial Benchmarks

Decoupling Spatial Processing from the Context Window. Generalist coding models stumble on planetary-scale analysis because multi-band satellite data and geospatial boundaries quickly saturate context limits. Google's Geospatial Reasoning agent solves this by introducing an upfront planning phase combined with an isolated execution workspace. The agent first constructs an analytical plan, checks dataset coverage, and processes heavy GeoJSON or raster operations outside the prompt context, returning only synthesised tabular and vector outputs. ### Integration via the Agent-to-Agent API. Google is expanding developer access via a new Agent-to-Agent (A2A) interface that exposes modular skills directly to custom Gemini workflows. Developers can join private organizational data—provided as CSV files, GeoJSON boundaries, or cloud tables—with Earth AI models including WeatherNext, Flood Forecasting, and the Population Dynamics Foundation Model (PDFM). In benchmark testing, fusing satellite embeddings with population signals yielded a 12% gain in health outcome predictions over standard baselines while generating maps at 8X finer spatial resolution. ### Open Benchmarks: GeoKTools and GeoExpertQueries. To evaluate geospatial reasoning systems systematically, Google published development splits for two new benchmarks. GeoKTools measures deterministic tool selection and execution across queries requiring between 1 and 4 distinct spatial capabilities. GeoExpertQueries evaluates multi-step, open-ended reasoning against expert-authored workflows across crisis resilience and public health. Benchmark evaluations showed that while standalone LLMs scored 93.6% on the web-grounded UnivEARTH benchmark, domain-specific tooling and dedicated planning phases were required to avoid failure on complex spatial computations.