Google Geospatial Reasoning Agent Moves Earth AI From Maps to Decisions

Google released its geospatial reasoning agent after four benchmark evaluations, moving Earth AI from isolated models toward automated, multi-step spatial analysis.

The experimental system connects Gemini with specialized models for weather, floods, population patterns, and remote sensing. It also searches large data catalogs, writes geospatial code, and trains task-specific predictive models. Google is expanding early access through Ask Google Earth, cloud workflows, and an Agent-to-Agent service for developers.

That combination creates the real tension behind the Google geospatial reasoning agent. The product is not competing mainly with another chatbot. It challenges the established workflow in which geographic information system specialists manually select datasets, reconcile formats, write code, and validate every analytical step.

Google says early partners have already used the agent for disaster preparation, disease investigations, clean-energy screening, and marketing analysis. Yet its own evaluation stops short of treating the output as a final answer. For open-ended tasks, the system is best understood as a first draft that experts must inspect.

The result is more consequential than another natural-language interface for maps. Google is trying to turn its collection of Earth models into an operating layer for decisions involving real places, people, and risks.

Google Geospatial Reasoning Agent Adds Planning to Earth AI

Google’s central change is orchestration: one agent can plan and execute workflows across models, datasets, maps, and code.

Google introduced Earth AI as a family of geospatial foundation models in 2025. A foundation model learns reusable representations that support many related tasks. The collection covered remote sensing, population dynamics, weather, floods, and other physical or social signals.

Those components could already produce useful predictions. However, a real planning question rarely fits within one model. An emergency team might need a storm forecast, administrative boundaries, population estimates, roads, and the locations of critical facilities.

The new geospatial agent is designed to connect those pieces. It starts with a dedicated planning stage, identifies required operations, and then selects appropriate tools and datasets.

Its available resources include WeatherNext, Flood Forecasting, Population Dynamics Foundations, AlphaEarth Foundations, and Remote Sensing Foundations. The agent can also reach Google Earth Engine, Data Commons, Google Maps Platform, and Dataset Search.

These sources do not fit neatly inside a language model’s context window. Satellite archives contain massive rasters, while map platforms expose routes, buildings, solar estimates, and air-quality information. Statistical repositories use different geographic units and reporting schedules.

Google addresses that mismatch with modular skills and a dedicated spatial workspace. A skill packages a specialized model or operation so the agent can invoke it only when needed. The workspace handles map layers and intermediate results without filling the conversation with raw spatial data.

The planning stage also asks what the system cannot do. It considers dataset availability, geographic coverage, required transformations, and possible fallback sources before launching an analysis. That design aims to reduce unsupported assumptions and shortcuts.

This is an important distinction from simply asking Gemini a geographic question. A general model can retrieve text about a location or generate code. It does not automatically know which satellite product, coordinate system, time range, or spatial operation fits an expert task.

The Google Earth AI agent instead builds a workflow. It can discover data, prepare features, run spatial intersections, create predictive models, compare their performance, and preserve data attribution.

Google is also widening access through three routes. Ask Google Earth offers an experimental deep-research experience. Cloud workflows support organization-specific analysis, while an Agent-to-Agent interface lets developers connect Earth AI with Gemini applications.

Access remains limited and experimental. Google has opened an interest process rather than announcing unrestricted availability, standard service guarantees, or production-scale adoption.

That limitation matters because the announcement represents a transition, not a finished platform. Earth AI now has an agentic execution layer, but Google still controls access to the most consequential capabilities.

Why Domain Tools Beat a General-Purpose Chatbot

The evaluations suggest that specialized tools create the largest improvement, while explicit planning matters most on open-ended expert work.

Google’s technical report compares four system configurations using Gemini 3.8 Flash and Gemini 3.1 Pro. The configurations isolate the value of general tools, Earth AI skills, and upfront planning.

The first configuration uses a standalone language model. The second adds common tools such as web search, map grounding, and code execution. The third supplies Earth AI skills but removes the dedicated planning and execution guidance.

The final configuration is the complete Google geospatial reasoning agent. It combines specialized skills with a planning stage that decomposes each request into a sequence of operations.

Google evaluates these configurations across four benchmarks. GeoKTools tests whether an agent can select and combine specific geospatial tools. GeoExpertQueries focuses on open-ended workflows derived from practitioner problems.

The study also uses CBGB and UnivEARTH, two external benchmarks. UnivEARTH appears saturated because even the standalone language model scores 93.6 percent. Google suggests public NASA articles may already expose the relevant questions and answers online.

That result is a useful warning about benchmark interpretation. A high score can reflect web retrieval or training-data familiarity rather than dependable spatial analysis.

On the other evaluations, basic web, map, and coding tools struggle with quantitative spatial tasks. Adding Earth AI skills produces the largest improvement across GeoExpertQueries, GeoKTools, and CBGB, according to Google.

Planning creates a further distinction. Specialized tools are sufficient for narrower problems with clearer paths. Open-ended GeoExpertQueries tasks benefit more from planning because the agent must choose datasets, sequence operations, and avoid convenient substitutes.

This mechanism explains how geospatial agents work in Google’s architecture. The language model does not replace weather science, hydrology, remote sensing, or demographic modeling. It coordinates systems built for those domains.

That coordination can also change the economics of model selection. The report says planning and execution guidance allowed Gemini 3.8 Flash to match or exceed Gemini 3.1 Pro on open-ended problems at lower cost.

The claim does not mean a smaller model becomes universally superior. It indicates that better tools and orchestration can matter more than model scale for structured professional workflows.

Google also released development portions of GeoKTools and GeoExpertQueries. The public benchmark materials include reference workflows, scoring resources, and evaluation tools for other researchers.

That release gives outsiders a way to test competing agents under similar protocols. However, development sets do not expose every item used in the reported evaluation. Holding back evaluation data helps limit direct benchmark optimization.

The strongest lesson is architectural. General reasoning alone does not solve specialized spatial work. An agent needs domain models, trustworthy datasets, deterministic operations, and a plan that connects them.

The result also pressures vendors that treat a chatbot interface as the primary innovation. Natural language lowers the entry barrier, but reliable spatial analysis depends on what the system invokes after understanding the request.

The Real Opponent Is Manual GIS Coordination

Google is compressing the coordination around geographic analysis, not eliminating the expertise required to judge it.

Traditional geographic information system workflows distribute responsibility across several roles. Subject experts frame the question, GIS analysts find compatible datasets, data scientists build models, and decision-makers interpret the output.

The process can take days because spatial data is unusually fragmented. A boundary file may use one coordinate reference system, while a satellite product uses another. Demographic statistics may describe counties when the operational question concerns villages or postal codes.

Time also complicates the analysis. A current storm forecast must align with older population estimates, recent infrastructure data, and historical exposure. Mixing those periods carelessly can create a polished but misleading map.

The Google geospatial reasoning agent tries to absorb much of that coordination. Its plan can specify the data sources, transformations, analytical steps, and model checks before execution begins.

Google’s earlier Earth AI research illustrated the concept through hurricane vulnerability analysis. The system combined predicted wind zones, population information, county boundaries, and satellite-based infrastructure detection.

That earlier evaluation reported 0.82 accuracy on its geospatial question-and-answer benchmark. Gemini 2.5 Pro scored 0.50, while Gemini 2.5 Flash scored 0.39.

The latest work moves beyond that earlier question-and-answer framing. The agent can now support dataset discovery, feature engineering, model training, performance comparisons, and interactive review.

Partner trials show what workflow compression looks like outside a benchmark. People in Need asked the system to analyze 14 historical cyclone paths and identify villages with high exposure.

According to the organization’s GIS coordinator, prepared data that previously required days of manual work produced results within minutes. That account is a partner statement, not an independent timing study.

GiveDirectly used custom boundaries to assess populations exposed to different wind-speed zones during Tropical Cyclone Maila. It also used the workspace to examine tornado damage in Mississippi.

The Milken Institute supplied more than 1,500 candidate buildings for potential community resilience hubs. The workflow combined rooftop solar potential, mobility information, nearby transit, climate risk, and federal investment-zone eligibility.

WPP connected Population Dynamics Insights with proprietary information for geospatial marketing analysis. Google says the workflow preserved the privacy of WPP’s data, although the announcement provides no independent privacy audit.

These cases reveal the main competitive boundary. The agent is not replacing one map product with another. It is competing with the handoffs, scripts, searches, and judgment calls surrounding existing spatial systems.

That puts established GIS workflows under pressure, including processes built around Esri, open-source spatial software, cloud warehouses, and custom notebooks. Google does not need to displace those components to capture part of their coordination layer.

The agent could instead become the interface that selects tools and prepares first-pass analysis. Existing systems may remain underneath, while professionals spend more time validating assumptions and less time assembling routine steps.

This shift resembles other knowledge-work agents. The main value comes from connecting specialized resources and maintaining context across a long task. Teams building similar systems can benefit from a searchable engineering knowledge base for datasets, assumptions, and review decisions.

However, geospatial work raises the stakes. A weak summary can waste a meeting. A weak exposure map can redirect emergency resources, exclude a community, or misstate infrastructure risk.

The manual workflow therefore remains the primary opponent, but expert oversight remains part of the product. Google’s own evidence supports acceleration more clearly than autonomy.

Partner Trials Put Earth AI Into Operational Decisions

The partner tests matter because they move Google Earth AI from model demonstrations into decisions with geographic and human consequences.

The World Health Organization Regional Office for Africa explored workflows for disease outbreak investigation. Such work can require environmental conditions, population patterns, health records, and administrative boundaries.

Google has not published a controlled outcome study showing improved public-health decisions. The trial demonstrates workflow feasibility, not verified epidemiological accuracy.

People in Need focused on anticipatory action around cyclones. Historical storm tracks can help identify settlements that repeatedly face damaging wind or rainfall. Yet exposure estimates still depend on boundary quality and current population information.

GiveDirectly’s use case brings another operational constraint. Humanitarian organizations often work with custom service areas that do not match official administrative boundaries. Uploading those areas allows the analysis to reflect an organization’s actual deployment plan.

The Milken Institute example shows the same approach in infrastructure planning. A resilience hub requires more than an available roof. Analysts must consider solar potential, transit access, climate hazards, mobility patterns, and policy eligibility.

An agent can join those layers quickly, but it also inherits every limitation within them. An outdated building list, incomplete transit dataset, or coarse climate layer can change the ranking.

Google describes its planetary prediction capabilities as another step in the same direction. The agent can discover data, construct features, train models, and compare candidate approaches for a user-defined prediction task.

That workflow moves beyond retrieving existing forecasts. It allows an organization to combine its own labels with Earth AI representations and create a customized model.

Google previously reported that combining Population Dynamics Foundations with AlphaEarth Foundations improved predictions of FEMA’s National Risk Index. Average R² increased by 11 percent across 20 hazards compared with using either representation alone.

The largest reported improvements were 25 percent for tornado risk and 17 percent for riverine flood risk. These measurements concerned prediction of an existing risk index, not direct prevention of disasters.

Google’s broader Earth AI portfolio already supports services at substantial scale. Its flood forecasting systems provide riverine warnings as far as seven days ahead across basins in more than 100 countries.

That operational history gives Google assets that a new geospatial-agent company would struggle to reproduce. It controls models, cloud infrastructure, mapping products, and large collections of geographic information.

It also creates concentration risk. Organizations could become dependent on one provider for data discovery, model representations, analysis, and deployment.

Open standards can reduce some of that dependency. Google’s Agent-to-Agent approach lets another application request Earth AI capabilities without copying the entire workflow into one interface.

The practical question is how much control developers receive. Early-access materials describe customization, private data uploads, interactive dataset review, and model evaluation. They do not yet define production quotas, regional availability, or long-term portability.

This uncertainty separates an impressive trial from a dependable operating system. Public agencies and humanitarian groups require stable access, documented provenance, repeatable outputs, and procedures for challenging a result.

Commercial organizations need those controls too. Site selection, infrastructure investment, insurance, and marketing decisions can affect communities even when public safety is not the stated objective.

Google’s announcement therefore expands the audience for geospatial AI. It also expands the number of people who must understand the limits of spatial evidence.

What the Benchmark Cannot Validate

The largest unresolved issue is not whether the agent can complete a workflow, but whether its answer is correct enough for the decision at hand.

GeoExpertQueries evaluates open-ended work by comparing an agent’s process with workflows developed by geospatial experts. Three independent experts contribute to the consensus reference, and a validated automated rater scores the result.

This approach addresses a difficult problem. Many spatial questions have several defensible analytical paths, and changing real-world conditions can make one permanent answer impossible.

However, workflow agreement is not the same as final-answer accuracy. Google explicitly says the resulting analysis should be treated as a strong first draft for experts to review.

The distinction deserves emphasis. An agent can choose reasonable datasets and operations while still using stale inputs, overlooking local context, or producing an incorrect output.

GeoKTools offers deterministic reference solvers, which make results easier to verify. Those templated tasks cannot capture every ambiguity found in public-health, humanitarian, or infrastructure decisions.

The benchmark suite also comes primarily from the system’s developer. Google created GeoExpertQueries and GeoKTools, selected the configurations, and reported the comparisons.

Releasing development data and scoring tools improves transparency. Independent replication across additional regions, models, and problem types remains necessary.

Geographic representation presents another challenge. Satellite coverage is global, but labels, administrative data, and ground observations are not equally reliable everywhere.

An independent benchmark found that geospatial foundation models do not consistently outperform supervised models. Its authors also identified geographic bias toward North America and Europe in common evaluation practices.

That research does not test Google’s latest complete agent. It does challenge the assumption that a general geospatial representation automatically transfers well across sensors, tasks, and regions.

Coverage gaps can produce particularly harmful errors. A fallback dataset may be technically available but too coarse for village-level planning. A population estimate can miss recent displacement after a conflict or disaster.

Spatial predictions can also conceal uncertainty behind attractive map layers. Colors and boundaries look precise even when the underlying inputs have wide error ranges.

A credible production system needs to show source dates, resolution, geographic coverage, transformations, model performance, and uncertainty. Reviewers should also be able to reproduce the decisive operations outside the agent.

Google says the workspace preserves attribution and supports transparent reasoning. Public materials do not yet establish whether every partner workflow meets a common audit standard.

Privacy requires similar scrutiny. Uploading facility locations, health data, customer information, or operational boundaries can reveal sensitive patterns even without personal names.

Google says WPP connected proprietary information while preserving data privacy. The available case summary does not specify retention controls, isolation guarantees, or independent testing.

Automation adds a final risk. The system can train custom predictive models, which means errors can move from one analysis into repeated future decisions.

Human review must therefore examine both the immediate output and the model-development process. Teams need clear approval gates for data selection, feature construction, evaluation, and deployment.

The Google geospatial reasoning agent makes those stages easier to initiate. It does not remove responsibility for defending them.

Three Signals Will Show Whether Geospatial Agents Are Ready

The next test is whether Google can turn promising, supervised experiments into reproducible workflows that outside experts trust.

The first signal is independent benchmark replication. Researchers should run the released GeoKTools and GeoExpertQueries materials against other models, tool stacks, and orchestration systems.

Results across regions will matter more than one aggregate score. Strong performance in data-rich markets would not establish dependable behavior in places with sparse labels or outdated boundaries.

Replication could strengthen Google’s case if different evaluators observe the same benefits from domain tools and upfront planning. Large variations would weaken claims about general planetary-scale reasoning.

The second signal is broader access with operational controls. Ask Google Earth, cloud workflows, and the Agent-to-Agent service are currently limited experimental routes.

Production readiness would require clear documentation for data provenance, reproducibility, permissions, private-data handling, error reporting, and model monitoring. Developers will also need stable interfaces and predictable geographic coverage.

These controls would show that Earth AI is becoming an infrastructure product rather than a collection of demonstrations. Continued invitation-only trials would keep the evidence narrow.

The third signal is measurable partner adoption. The current examples describe time savings and useful analyses, but they do not provide controlled measurements of decision quality.

Future case studies should report whether organizations repeated the workflows, how often experts corrected the outputs, and whether results changed operational decisions.

For disaster response, useful measures include analysis time, missed communities, false alerts, and agreement with validated field information. Public-health deployments require disease-specific validation and local governance.

Infrastructure users should disclose how the agent’s rankings compare with expert assessments. Commercial users should explain whether proprietary data improved results without creating unacceptable privacy exposure.

Google also faces a strategic test from the wider GIS market. Mapping providers, satellite companies, cloud platforms, and open-source projects can connect language models to their own spatial tools.

Their response will determine whether Google owns the orchestration layer or helps establish a broader category. Open benchmarks make that comparison easier, while Google’s integrated data and models create a substantial advantage.

For developers, the immediate opportunity is not to remove geospatial specialists. It is to shorten data discovery, automate repeatable operations, and preserve a reviewable record of every decision.

For enterprise buyers, the key question is equally concrete: can the agent reproduce a result when the data changes, an expert challenges an assumption, or a regulator requests evidence?

The Google Earth AI agent deserves attention because it joins models, tools, and planning within one workflow. Its value will depend on transparent evaluation rather than the fluency of its interface.

Teams evaluating the system should select one bounded problem, define an expert baseline, and record every correction. They should test weak-data regions, not only familiar examples.

They should also decide which outputs can inform a human and which can trigger an action. That boundary should exist before deployment, not after the first serious error.

The Google geospatial reasoning agent has made planetary analysis faster to assemble. The next three signals will show whether it can also make that analysis dependable enough to use.