AgenticGEO: AI for Geospatial Workflows

Updated 5 July 2026

AgenticGEO is a framework for building AI agents that update geospatial state using structured, tool-mediated workflows.
It emphasizes geospatial, temporal, and physical validity by integrating CRS, resolution, and temporal variables into its state formalism.
The approach underpins applications in EO such as disaster response, urban planning, and data discovery through layered architectural designs.

Searching arXiv for the cited work to ground the synthesis in current papers. AgenticGEO (Editor’s term) denotes agentic AI for geospatial and Earth observation workflows: systems that interpret natural-language goals, operate over georeferenced and temporally structured data, invoke GIS and EO tools, update intermediate geospatial state, and are evaluated not only by final-answer accuracy but also by geospatial, temporal, and physical validity. The term is not formalized under that exact name in the primary position paper, but the paper explicitly develops the conceptual, formal, and architectural foundations for “EO-native agents,” “geospatial agents,” and “agentic EO models” organized around geospatial state, tool effects, and validity (Munir et al., 27 Apr 2026). A broader survey places this shift within the transition from static deep learning models to autonomous agentic AI in remote sensing (Talemi et al., 5 Jan 2026).

1. Conceptual definition and formal basis

A central reframing is that an EO agent is a geospatial state updater. Rather than treating tools as black-box functions mapping text to text, an EO-native agent operates over structured state

$s_t = (x_t, c_t, r_t, e_t, \tau_t, m_t, u_t, p_t, h_t),$

where $x_t$ is current data, $c_t$ is CRS, $r_t$ is spatial resolution or GSD, $e_t$ is spatial extent or footprint, $\tau_t$ is temporal window or coverage, $m_t$ is modality, $u_t$ is uncertainty or reliability, $p_t$ is provenance, and $h_t$ is tool-call history. An action is a parameterized tool invocation,

$x_t$ 0

and execution induces

$x_t$ 1

This formalization makes CRS, resolution, extent, modality, and time first-class state variables, and it makes order dependence and irreversibility explicit (Munir et al., 27 Apr 2026).

A complementary formalization appears in geo-analytical question answering, where natural-language questions are parsed into executable workflows represented as GeoFlow Graphs. There, geospatial reasoning is formulated as a concept transformation problem $x_t$ 2, with nodes labeled by core spatial concepts and edges representing transformations. Spatial-Agent defines

$x_t$ 3

with concept labels $x_t$ 4 and functional-role assignments $x_t$ 5, and constrains well-formed workflows through acyclicity, role ordering, type compatibility, data availability, and connectivity (Bao et al., 23 Jan 2026). This suggests that AgenticGEO has at least two compatible formal views: structured geospatial state transition and concept-transformation workflow graphs.

The same literature emphasizes that correctness in geospatial workflows is externally grounded. Internal logical coherence is insufficient when correctness depends on geospatial consistency, temporally valid comparisons, and physical validity. For this reason, the design space of AgenticGEO differs from generic agentic AI not by application-specific prompt engineering alone, but by a different operational semantics for state, tools, and verification (Munir et al., 27 Apr 2026).

2. Structural constraints and characteristic failure modes

EO workflows operate over georeferenced, multi-modal, and temporally structured data. Operations such as reprojection, resampling, clipping, mosaicking, compositing, aggregation, tiling, temporal differencing, and spectral index computation transform the underlying state and can constrain subsequent analysis. Because of this, errors can propagate silently across steps, and the final output may remain numerically plausible while being scientifically invalid (Munir et al., 27 Apr 2026).

Three structural constraints recur. First, geospatial consistency requires that tools which combine or compare layers respect CRS, grid alignment, extent, and units. A flood-map raster in EPSG:4326 and a damage map in UTM may still be subtracted pixel-wise without failing loudly, yet the result is meaningless. Second, multi-modal and multi-resolution consistency matters because EO workflows mix optical imagery, SAR backscatter, multispectral imagery, reanalysis grids, and vector layers with distinct resolutions, extents, and physical semantics. Third, temporal validity matters because EO time series are irregular and incomplete, with cloud gaps, seasonality, phenology, and multi-date comparisons that can mask or create spurious changes (Munir et al., 27 Apr 2026).

The position paper identifies several generic-agent assumptions that break in EO. Tools are not independent stateless functions; many actions are not reversible or are lossy; errors are often latent rather than locally detectable; and internal evaluation is not sufficient because correctness is defined by physics, geodesy, and observed Earth processes. To formalize this, the paper defines hard feasibility predicates

$x_t$ 6

with

$x_t$ 7

and a soft geospatial-quality score

$x_t$ 8

It also introduces an irreversibility penalty within step reward to discourage transitions that unnecessarily reduce future analytical flexibility (Munir et al., 27 Apr 2026).

The resulting failure modes are characteristic: geospatial inconsistency, temporal invalidity, physical implausibility, silent error propagation, and misuse of models outside their training domain. AgenticGEO, in this sense, is not merely a geospatial deployment of a generic LLM agent; it is an attempt to build agents whose reasoning and execution are constrained by geospatial and physical structure (Munir et al., 27 Apr 2026).

3. Architectural patterns and agency primitives

A recurring architectural pattern is Planner–Executor–Verifier (PEV). In the EO-native formulation, the planner designs a workflow given a task and initial state, the executor executes tools with precise arguments and updates $x_t$ 9, and the verifier checks each transition for geometric consistency, temporal validity, physical plausibility, provenance consistency, and statistical reliability, using

$c_t$ 0

normalized to $c_t$ 1 (Munir et al., 27 Apr 2026).

A parallel line of work defines agency primitives for GeoAI assistants: navigation, perception, geo-referenced memory, Earth embeddings, compute graphs, budgets, propagation, attribution, and dual modeling. These primitives are explicitly framed as capabilities that let an assistant take grounded actions in a GIS workspace while keeping the user in control, thereby connecting foundation models to artifact-centric, human-in-the-loop GIS workflows (Zaytar et al., 2 Apr 2026).

Representative systems illustrate how these abstractions are instantiated.

System	Architectural emphasis	Paper
OpenEarthAgent	Unified tool registry, working memory, tool-augmented trajectories	(Shabbir et al., 19 Feb 2026)
Spatial-Agent	GeoFlow Graphs, concept transformation, typed workflow constraints	(Bao et al., 23 Jan 2026)
MapAgent	Hierarchical planner plus dedicated map-tool agent	(Hasan et al., 7 Sep 2025)
GeoJSON Agents	Planner–Worker multi-agent execution over GeoJSON	(Luo et al., 10 Sep 2025)
City Editing (CEAE)	Hierarchical planner, geoexecutors, validator, aggregator	(Liu et al., 22 Feb 2026)

OpenEarthAgent implements a tool-augmented multimodal agent with a unified tool schema $c_t$ 2, working memory over instructions, imagery, past tool calls, and spatial metadata, and an execution environment that replays and validates trajectories before training (Shabbir et al., 19 Feb 2026). MapAgent separates a top-level planner from a map-service module with a dedicated map-tool agent and four higher-level tools—Trip, Route, Nearby, and PlaceInfo—specifically to reduce tool inflation and improve coordination across similar APIs (Hasan et al., 7 Sep 2025). GeoJSON Agents use a Planner agent that interprets natural-language tasks into structured GeoJSON commands and Worker agents that either invoke predefined function APIs or dynamically generate and execute Python-based spatial analysis code (Luo et al., 10 Sep 2025). City Editing formulates urban renewal as machine-executable editing over GeoJSON, decomposed into polygon-, line-, and point-level intents, with explicit propagation of intermediate spatial constraints and an iterative execution-validation loop (Liu et al., 22 Feb 2026).

Across these systems, AgenticGEO appears less as a single architecture than as a family of architectures sharing several properties: explicit spatial state or workflow representation, typed tool use, intermediate artifact management, and some form of validation or human review before commitment.

4. Learning paradigms, benchmarks, and evaluation

The literature repeatedly argues that standard benchmarks are insufficient because they focus on static inputs, one-step outputs, or simplistic tools. AgenticGEO therefore shifts attention toward trajectory supervision, tool-use fidelity, workflow correctness, and state consistency (Munir et al., 27 Apr 2026).

OpenEarthAgent exemplifies supervised trajectory learning. Its corpus comprises 14,538 training and 1,169 evaluation instances, with 100,656 steps in training and 7,064 steps in evaluation, spanning urban, environmental, disaster, and infrastructure domains and incorporating GIS-based operations alongside NDVI, NBR, and NDBI analyses. It is trained with a maximum-likelihood objective over tool actions,

$c_t$ 3

and reports, in tool-agnostic mode, improvements from a baseline Qwen3-4B to OpenEarthAgent-4B from Inst. 97.34 to 99.51, Tool. 86.94 to 97.18, ArgN. 84.12 to 96.08, and ArgV. 33.55 to 62.10. In end-to-end evaluation it reaches Per. 58.30, Op. 56.76, Logic. 51.18, GIS. 98.52, AnyOrder 67.75, SameOrder 67.24, Unique 72.71, and Ans. 45.26 (Shabbir et al., 19 Feb 2026).

The EO-native position paper proposes trajectory-level evaluation metrics intended to complement final-answer metrics such as IoU and RMSE. These include Pipeline Integrity (PI),

$c_t$ 4

Trajectory Validity Score (TVS),

$c_t$ 5

Discounted Inconsistency Burden (DIB),

$c_t$ 6

and cost-aware efficiency,

$c_t$ 7

The same paper advocates hybrid SFT + RL, constrained self-improvement, and verifier-guided rewards aligned with geospatial and physical validity (Munir et al., 27 Apr 2026).

GeoDisaster extends this logic to operational disaster analysis. It contains 2,921 verified instances across 43 question types and five task families, with 18 disaster-oriented tools and an orchestrated multi-agent framework aligned by Role-Contract Expectation Alignment (RCEA), described as failure-aware supervised fine-tuning combined with contract-grounded reinforcement learning over dense step-level signals (Hasan et al., 15 Jun 2026). GeoVista addresses geolocalization with GeoBench, whose test set contains 1,142 images—512 standard photos, 512 panoramas, and 108 satellite images—and trains GeoVista with cold-start SFT followed by GRPO with a hierarchical geographic reward. GeoVista-7B reaches Country 92.64, Province 79.60, City 72.68, with Panorama city 79.49, Photo 72.27, Satellite 44.92, and nuanced performance of 52.83% within 3 km and median distance 2.35 km (Wang et al., 19 Nov 2025).

For data discovery rather than mapping or diagnosis, NASA-EO-Bench contributes 47,654 query–dataset positive pairs and 21,272 task-based queries over 8,058 CMR datasets. On this benchmark, NN-SSC + BM25 (Hybrid) reaches $c_t$ 8 and $c_t$ 9, while agentic reranking on top of the strong retriever lifts MRR from 0.302 to 0.388 for Opus 4.7 on a stratified subset (Yu et al., 2 Jul 2026). Taken together, these benchmarks make clear that AgenticGEO is evaluated increasingly as workflow execution under spatial constraints rather than as isolated prediction.

5. Representative systems and operational domains

AgenticGEO spans multiple operational domains rather than a single canonical task. Disaster response is a dominant setting: GeoDisaster covers deforestation monitoring, multi-hazard analysis, building-damage assessment, flood-safe routing, and Sentinel-1 SAR flood monitoring, explicitly connecting hazard detection, exposure estimation, routing, and diagnostic report generation within executable tool chains (Hasan et al., 15 Jun 2026). GeoVista addresses geolocalization through web-augmented agentic visual reasoning with an image crop-and-zoom tool and a web-search tool, showing how high-resolution imagery and external information retrieval can be integrated within a thought–action–observation loop (Wang et al., 19 Nov 2025).

In geospatial data discovery, an agentic search system has been deployed as a public service for the geoscience community, taking natural-language research queries and returning matching NASA datasets and tools. Its architecture combines lexical retrieval, task-adapted neural scoring, and an agentic reranking stage that can call external tools such as arXiv and web search (Yu et al., 2 Jul 2026). In map-based geospatial reasoning, MapAgent targets route planning, POI search, and multi-hop reasoning over live map services, using a hierarchical planner and a dedicated map-tool agent to orchestrate related APIs adaptively in parallel (Hasan et al., 7 Sep 2025).

Urban and planning workflows form another cluster. City Editing treats urban renewal as iterative modification of existing urban plans represented in GeoJSON, with hierarchical geometric intents over polygon-, line-, and point-level operations and explicit propagation of intermediate spatial constraints (Liu et al., 22 Feb 2026). GeoJSON Agents likewise operate over GeoJSON as the canonical state representation for GIS automation and show that the Function Calling-based GeoJSON Agent achieved an accuracy of 85.71%, while the Code Generation-based agent reached 97.14%, both outperforming the best-performing general-purpose model at 48.57% (Luo et al., 10 Sep 2025).

Scientific and industrial specializations also appear. GeoSR frames geospatial prediction as an iterative self-refinement loop grounded in Tobler’s First Law of Geography, with a variable-selection agent, a point-selection agent, and a refine agent; it reports, for example, GPT-3.5-Turbo improvement on Infant Mortality from 0.445 to 0.747 and bias from -0.188 to -0.006 (Tang et al., 6 Aug 2025). GAIA, in geothermal field development, combines GAIA Agent, GAIA Chat, and GAIA Digital Twin in an agentic retrieval-augmented workflow spanning knowledge-base retrieval, phase picking, event location, magnitude estimation, seismicity forecasting, and interactive visualization (Harsuko et al., 5 Nov 2025).

A broader infrastructural implication appears in “The Agentic Economy,” which distinguishes assistant agents and service agents and asks whether inter-agent communication will occur within closed “agentic walled gardens” or through a more open “web of agents” (Rothschild et al., 21 May 2025). Applied to AgenticGEO, this suggests that geospatial agents may ultimately form not only analytical systems but also interoperable ecosystems linking citizen assistants, city-service agents, EO platforms, digital twins, and regulatory services.

6. Alignment, security, and unresolved research directions

Because geospatial outputs can be regionally contingent, AgenticGEO raises an alignment problem that is explicitly spatial. “Whose Truth? Pluralistic Geo-Alignment for (Agentic) AI” defines geo-alignment in terms of matching the system distribution $r_t$ 0 to the locally appropriate distribution $r_t$ 1 under geographic context $r_t$ 2: $r_t$ 3 The paper emphasizes that what is truthful, appropriate, or legal can differ across regions, and that agentic systems need spatio-temporally aware alignment rather than one-size-fits-all approaches (Janowicz et al., 7 Aug 2025). For AgenticGEO, this matters not only for answers but also for actions: where to go, what laws apply, which safety policies or contested border representations to use.

Security is equally central. A security-oriented framework for multi-agent GIS systems evaluates prompt injection, unauthorized access, off-topic drift, competitor probing, vague-location abuse, and format sabotage using a red-teaming framework with an adaptive attacker LLM and a deterministic judge. In reported results, SEC-1 and SEC-2 remain at 0% attacker success before and after optimization; TOP-1 drops from 84% to 55%; REC-1 drops from 10% to 0%; REC-3 drops from 33% to 20%; while TOP-2 rises from 0% to 13%, illustrating that prompt optimization can improve some attack surfaces while degrading others (Gao et al., 13 Jun 2026). This is a reminder that AgenticGEO systems require not just prompt engineering but explicit state-machine orchestration, least-privilege tool access, logging, and possibly LLM-as-judge safeguards for GIS queries and API calls.

Open problems remain structural. The EO-native research agenda calls for better state representations and uncertainty modeling, tool schema learning, robustness to cloud masks and missing data, stronger verifier modules, more realistic benchmarks, and constrained self-improvement gated by verifier scores (Munir et al., 27 Apr 2026). The survey literature adds limitations in grounding, safety, orchestration, shallow temporal memory, fragmented evaluation, and the need for EO-native perception backbones, hierarchical memory, and full-path evaluation beyond final outputs (Talemi et al., 5 Jan 2026). In that sense, AgenticGEO is best understood not as a finished architecture but as an emerging class of geospatially grounded agentic systems whose defining problem is to make planning, tool use, memory, validation, and alignment cohere under the physical, spatial, and workflow constraints that govern GIS and Earth observation (Munir et al., 27 Apr 2026).