Google's Geospatial Reasoning agent starts every query with a planning stage
Google Research is widening experimental access to its Geospatial Reasoning agent, which its report tests in four configurations on two Gemini models. Access runs through Ask Google Earth and an A2A service.
The Engineer · Build desk
What happened
- Google packages Earth AI capabilities as modular skills that the agent calls only when a user question needs them.
- Every response from the agent shows its plan and the data sources it used, so experts can check assumptions or substitute datasets.
- Users can bring their own data through the agent API as CSV files, GeoJSON boundaries or cloud tables and join it with Earth AI's global models.
Why it matters
- constraint The claim that standalone LLMs cannot do this work rests on Google's own technical report, and the post names only two of its four benchmarks as externally developed.
- decision Displayed plans turn dataset substitution into a review task, so a team adopting the agent has to decide who reads the plan and signs off on fallback sources.
- capability A team holding local outcome data such as crop yields or grid outages can ask the agent to select signals and train validated prediction models for it.
Start with what the agent keeps out of its context. Large map layers are processed in a dedicated workspace [5]. Google says this avoids overwhelming the agent context with massive spatial data while preserving accuracy and tracking data attribution [6]. By Google's account the problem is real: planetary datasets are massive, come in specialized map and satellite formats, and need careful multi-step planning to combine [17].
Planning comes first. The full agent begins every query by reasoning over its available capabilities to work out what it can and cannot do, then sets out the stages of retrieval and processing [12]. Before complex operations it also builds a step-by-step plan and picks fallback datasets where local coverage is sparse [4]. We would test the fallback step first. A wrong substitute changes the answer without changing its shape.
The report's comparison is built to separate causes. It runs four configurations on Gemini 3.8 Flash and Gemini 3.1 Pro: a bare LLM, the LLM with web search and maps grounding, the LLM with Earth AI tools but no upfront planning, and the full agent [11]. Compare the third with the fourth and the difference is planning. Compare the second with the third and the difference is the Earth AI tools. The design fits the claim. Shetty and Barrington wrote that the combination "enables a Geospatial Reasoning agent to carry out complex spatial analysis and planetary prediction" that standalone LLMs cannot do [16].
Transfer is the open question. Two of the four benchmarks, CBGB and UnivEARTH, were developed outside Google and test geospatial code generation and question answering [14]. Scores there should carry to a workload shaped like code generation or question answering over public data. The 12% figure comes from a different test: Google's public health research, which compared population and satellite signals with demographic baselines [10]. For it to transfer, a team would need local outcome data of the kind the Planetary Prediction Engine takes, such as disease rates, crop yields or grid outages [9], plus coverage from the underlying signals in its own region. The text of the post available to us breaks off before the benchmark scores, and it does not name the health outcome or region behind the 12%.
Access is still labelled experimental. Google describes the agent as an experimental system that grew out of an early concept [15], and it is widening access through Ask Google Earth and an A2A service after tests with partners [2].
What to watch
- The technical report's benchmark tables: the gap between the no-planning configuration and the full agent, on each of the two Gemini models.
- Partner results from Ask Google Earth or the A2A service in regions where the agent has to fall back to substitute datasets.
- Which health outcome and region sit behind the 12% gain and the 8X finer mapping.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption
- Insufficient
- Hype gap+20
- Incentives60
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The Google Research post by Shravya Shetty and Luke Barrington is dated October 14, 2026.
- [2]
Following promising tests with partners, Google is expanding experimental agent access via Ask Google Earth and an Agent-to-Agent (A2A) service.
- [3]
Earth AI capabilities are packaged as modular skills that the agent calls only when a user question needs them.
- [4]
Before running complex operations, the agent builds a step-by-step analytical plan and selects appropriate fallback datasets when local coverage is sparse.
- [5]
The agent processes large map layers in a dedicated workspace.
- [6]
Google says this approach avoids overwhelming the agent context with massive spatial data while preserving accuracy and tracking data attribution.
- [7]
Every response shows the agent's plan and the data sources used so experts can check assumptions or substitute datasets.
- [8]
Users can input their own datasets via the Geospatial Reasoning agent API as CSV spreadsheets, custom map boundaries (GeoJSON) or cloud tables, and ask the agent to join their private locations with Earth AI's global models and catalogs.
- [9]
Given a user's local outcome data, such as regional disease rates, crop yields or grid outages, the agent powering the Planetary Prediction Engine automatically selects relevant population and environmental signals, fuses Population Dynamics Insights and Custom Satellite Embeddings, and trains validated machine learning models for spatial prediction, super-resolution downscaling and forecasting.
- [10]
In Google's recent public health research, combining population and satellite signals improved predictions of health outcomes by 12% over demographic baselines and enabled mapping at 8X finer spatial resolution.
- [11]
The technical report compares four agent configurations across Gemini 3.8 Flash and Gemini 3.1 Pro: a standalone LLM; the LLM with basic web search and maps grounding tools; the LLM with Earth AI tools but without upfront planning; and the full Geospatial Reasoning agent.
- [12]
The full Geospatial Reasoning agent begins every query with a dedicated planning stage that reasons over available capabilities to assess what it can and cannot do and sets out the stages of retrieval and processing needed to reach an answer.
- [13]
The evaluation spans four benchmarks, covering expert-level open-ended analysis, multi-tool orchestration, and community-developed geospatial problems with deterministic answers.
- [14]
CBGB and UnivEARTH are two externally developed benchmarks that test geospatial code generation and question answering respectively.
- [15]
Google says it has evolved the Geospatial Reasoning agent from an early concept into a much more capable, experimental system built for multi-step geospatial reasoning across planet-scale data.
- [16]
Shetty and Barrington wrote that combining domain-specific foundation models, spatial tools and upfront planning "enables a Geospatial Reasoning agent to carry out complex spatial analysis and planetary prediction that standalone LLMs can't."
- [17]
Google says general-purpose LLMs and standard coding tools struggle with real-world spatial problems because planetary datasets are massive, with specialized map and satellite formats that require careful multi-step planning to combine accurately.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- research.googlePutting Earth AI to work: Geospatial agents for our planet's biggest challenges
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.