PANGAEA-GPT: Autonomous Geoscience Workflow

• PANGAEA-GPT is a hierarchical multi-agent framework for automating geoscientific data discovery, harmonizing heterogeneous metadata and formats.
• It employs a centralized supervisor and specialized worker agents to execute search, analysis, visualization, and narrative synthesis in a secure sandbox.
• Benchmark results show significant gains in retrieval precision and workflow automation, addressing data reuse challenges in large Earth science archives.

PANGAEA-GPT is a hierarchical multi-agent framework for autonomous data discovery and analysis in geoscientific repositories, developed especially around the PANGAEA archive. Its defining aim is to turn a natural-language scientific question into an end-to-end workflow that can discover datasets, inspect heterogeneous metadata, generate and execute analysis code in a sandboxed environment, repair failures from execution feedback, produce figures, and synthesize a narrative report. In the geoscience setting, the system is positioned as a response to the mismatch between very large archival holdings and comparatively low reuse, and as an alternative to treating repository interaction as a simple search or plotting problem (Pantiukhin et al., 24 Feb 2026, Pantiukhin et al., 7 Mar 2025).

1. Problem setting and motivation

PANGAEA-GPT was motivated by a data-reuse problem rather than by a generic chatbot objective. The archive-centered paper starts from the observation that PANGAEA hosts over 400,000 curated datasets and 800+ parameters, while citation metrics indicate that a substantial portion remains underutilized; it cites evidence that nearly 90% of archived datasets have never been cited in the peer-reviewed literature. The earlier Perspective frames the same problem more broadly across Earth science repositories, noting that more than 93% of datasets are uncited and that archives contain millions of heterogeneous datasets and several petabytes of data (Pantiukhin et al., 24 Feb 2026, Pantiukhin et al., 7 Mar 2025).

The technical bottlenecks are described as metadata incompleteness or inconsistency, format heterogeneity, domain-specific reasoning burden, and workflow fragmentation. The same physical quantity may appear as a column in CSV/tabular data, a variable in NetCDF, a chunk in Zarr, or a product accessed through an external API. Users must often translate terms such as “Fram Strait,” “winter 2013,” or “surface temperature” into bounding boxes, seasonal constraints, product types, and coordinate conventions before any analysis can begin. PANGAEA-GPT was designed to absorb that translation burden and to connect discovery, harmonization, analysis, visualization, and reporting inside one coordinated pipeline (Pantiukhin et al., 24 Feb 2026).

A common misconception is that such a system is merely an LLM wrapper around search. The papers argue instead that the central difficulty lies in bridging natural-language intent to heterogeneous, poorly standardized, multi-format Earth-science data, and then executing correct analysis code safely and reproducibly. This places PANGAEA-GPT closer to an archive-native workflow engine than to a conventional conversational interface (Pantiukhin et al., 24 Feb 2026).

2. Hierarchical multi-agent design

The implemented system uses a centralized Supervisor-Worker topology with a clear separation between planning and execution. In the 2026 formulation, it is described as a stateful multi-agent system built as a directed cyclic graph with LangGraph on top of LangChain. The high-level flow is: a user asks a question in natural language; a Search Agent finds relevant datasets; the Supervisor Agent decomposes the task and routes subtasks; specialized worker agents perform analysis, visualization, and synthesis; and a Writer Agent assembles the final narrative report (Pantiukhin et al., 24 Feb 2026).

Centralized planning is the core organizational principle. The Supervisor maintains global conversation state, inspects dataset structure and metadata, decomposes user intent into subtasks, dispatches each subtask to an appropriate worker, and updates its plan as results arrive. The Perspective describes the broader MAS pipeline in similar terms: a supervisor agent acts as the command-and-control node, dynamically creates a task-specific agent subgraph, coordinates retrieval and tool usage, and consolidates outputs into a final response (Pantiukhin et al., 7 Mar 2025).

The specialist layer in the archive paper consists of five workers.

Agent	Primary scope	Main functions
Oceanographer Agent	NetCDF/Zarr and external climate APIs	Geospatial and spatiotemporal matchup, nearest-grid-point and 4D interpolation
Ecologist Agent	Biological and tabular datasets	Biodiversity metrics and taxonomic aggregation
Visualization Agent	Figures and plots	Plot generation with exemplar-plot RAG and reflexive QC
DataFrame Agent	Tabular manipulation only	Statistical analysis; explicitly prohibited from generating visualizations
Writer Agent	Narrative synthesis	Coherent prose and figure interpretation through a vision-language interface

This specialization is fundamental to the system’s robustness claims. PANGAEA-GPT is not presented as a single generalist agent that “does everything,” but as a planner that dispatches narrowly scoped work to specialists with restricted capabilities and task-appropriate tools (Pantiukhin et al., 24 Feb 2026).

3. Retrieval, metadata introspection, and routing logic

The search subsystem is evaluated in three configurations: a Baseline using Elasticsearch BM25 / tf-idf keyword matching; a Simple LLM that performs one-pass natural-language translation into a structured Boolean query; and Agentic Search, which uses a ReAct loop to infer constraints, execute a query, inspect results, generate multiple query permutations, and consolidate and rank results. This search stack is intended to resolve implicit scientific constraints that are often absent from the literal wording of the query (Pantiukhin et al., 24 Feb 2026).

A defining mechanism is strict data-type-aware routing. Rather than inferring data type only from user language, the system first introspects the selected dataset’s metadata. It pre-fetches parameter inventories, matrix dimensions, coordinate reference systems, temporal extents, file size, and indicators of whether the dataset includes geospatial coordinates, is a gridded array, has a depth axis, or has a temporal index. These properties are turned into binary feature flags that directly determine routing. The paper gives concrete examples: gridded array + depth axis routes to the Oceanographer Agent, while a flat table without spatial coordinates routes to the DataFrame Agent (Pantiukhin et al., 24 Feb 2026).

The purpose of this mechanism is to prevent category mistakes. It avoids sending NetCDF/Zarr arrays to tabular analysis code, trying to visualize a gridded geophysical field with table-only logic, or forcing a biological dataset through an ocean-model workflow. This suggests that the system’s main contribution is not only linguistic interpretation but metadata-conditioned workflow control (Pantiukhin et al., 24 Feb 2026).

The Perspective situates this routing logic within a broader MAS design that also uses retrieval-augmented generation, domain-specialized sandboxes, and validation agents. In that framing, PANGAEA-GPT is meant to automate the tedious, heterogeneous, and error-prone components of repository interaction while leaving scientific judgment outside the scope of pure orchestration (Pantiukhin et al., 7 Mar 2025).

4. Execution environment, self-correction, and quality control

PANGAEA-GPT executes generated Python code inside a sandboxed Jupyter environment. The execution kernel is persistent, isolated in a filesystem sandbox, and associated with a unique workspace identified by a UUID for each user session. This design provides state persistence, session isolation, and a direct channel for self-correction because standard error tracebacks are captured and fed back to the agent loop (Pantiukhin et al., 24 Feb 2026).

The repair cycle is explicit: generate code, execute, inspect output or traceback, revise the code, and retry. The archive paper gives several concrete error types that the system can repair through execution feedback: an invalid Copernicus surface-layer request retried with the nearest valid layer; timestamp mismatches between mooring data and model data; out-of-bounds interpolation for ship-track points; off-by-one binning errors in wind-rose generation; and division-by-zero in biodiversity calculations, resolved by masking zero-abundance taxa (Pantiukhin et al., 24 Feb 2026).

Visualization quality control is also agentic. The Visualization Agent has a retrieval-augmented index of exemplar scientific plots and a reflexive QC loop. Figures are critiqued by a vision model, and plots are revised if scores are too low. In one scenario, the critic detected that a legend obscured key plot elements and triggered a layout refactor. The broader Perspective generalizes this into a family of validation or reflection agents that can check whether depth axes are reversed correctly, whether unit conversions are necessary, whether metadata are suspicious, and whether variable names align with controlled vocabularies such as SeaDataNet (Pantiukhin et al., 7 Mar 2025).

The Perspective additionally describes a multi-tier memory architecture. Short-term memory holds active session context, while long-term memory is stored in a searchable vector database containing summaries, logs, and reusable outputs. A plausible implication is that PANGAEA-GPT is intended not merely for one-off query answering but for long-running exploratory sessions in which intermediate scientific context must survive beyond the immediate prompt window (Pantiukhin et al., 7 Mar 2025).

5. Benchmarks and scientific use cases

The search benchmark uses 100 natural language queries spanning six domains—physical oceanography, marine geology, biogeochemistry, paleoclimatology, marine ecology, and atmospheric science—and scores each query on Precision, Recall, Parameter Coverage, Metadata Quality, and Result Relevance using a 1–10 scale. Mean scores were 2.87/10 for the Baseline, 5.46/10 for the Simple LLM, and 8.14/10 for Agentic Search, with the largest gains in Precision at 8.53 and Parameter Coverage at 8.99. The paper also reports qualitative gains in seasonal reasoning and source disambiguation, such as mapping “winter 2013” in the Southern Hemisphere to June–August and prioritizing actual satellite altimetry products over in-situ datasets that only mention altimetry validation (Pantiukhin et al., 24 Feb 2026).

The archive paper validates the full pipeline with four end-to-end scenarios. In the Weddell Sea microplastic case, the system combined a PANGAEA microplastic dataset from 34 Manta trawl stations, DOI 10.1594/PANGAEA.941076, with Copernicus GLORYS12V1 daily surface velocity fields. It produced a map of current speed, streamlines, and stations scaled by microplastic abundance, visually capturing cyclonic rotation of the Weddell Gyre. A later analysis on 21 co-located stations found Spearman $\rho = -0.47$, $p = 0.033$, and Mann–Whitney $U = 88.0$, $p = 0.022$, with higher abundance in low-speed regions; the authors explicitly characterize this as exploratory rather than definitive ecological inference (Pantiukhin et al., 24 Feb 2026).

In the Fram Strait validation case, PANGAEA-GPT harmonized ten HAUSGARTEN mooring datasets, PANGAEA DOIs 845608–845617, and matched them against Copernicus GLORYS12V1 through a 4D oceanographic workflow. The system completed 135,678 co-locations autonomously and reported Bias $+0.35\,^\circ$C, RMSE $1.09\,^\circ$C, and Pearson $r = 0.31$. The paper interprets the warm bias as consistent with known GLORYS12V1 offsets in Fram Strait and the moderate correlation as partly attributable to nearest-grid-point matching in a steep gradient region (Pantiukhin et al., 24 Feb 2026).

In the resource-efficient trajectory analysis case, the system used the RV Polarstern MOSAiC PS122/1 navigation log, DOI 10.1594/PANGAEA.935221, together with ERA5 2-meter air temperature from Arraylake. It extracted spatiotemporal bounds from the navigation log, performed targeted cloud-native retrieval, interpolated ERA5 to 2,013 hourly ship positions, and generated a wind rose. The retrieval volume was minimized to about 260 MB by bounding-box selection, illustrating that the workflow can optimize resource use as well as automate alignment and plotting (Pantiukhin et al., 24 Feb 2026).

In the biodiversity–water mass coupling case, PANGAEA-GPT analyzed the North Atlantic jellyfish cruise dataset, DOI 10.1594/PANGAEA.829702, alongside Copernicus GLORYS12V1 temperature and salinity profiles. The Ecologist Agent split stations into an Eastern transect and a Western transect, computed Shannon-Wiener diversity index $H'$, and the Oceanographer Agent generated matched hydrographic profiles and a T-S diagram. The reported contrast was Welch’s $t = -3.39$, $p = 0.0014$, with $p = 0.033$0 and $p = 0.033$1. For the dominant hydrozoan Aglantha digitale, highest abundance clustered in cold, fresh water with $p = 0.033$2C and $p = 0.033$3–$p = 0.033$4 psu, described as consistent with Subpolar Mode Water / Irminger Current (Pantiukhin et al., 24 Feb 2026).

6. Limitations, misconceptions, and research trajectory

The system’s limitations are stated directly. The Supervisor is a single point of failure: if it misroutes a task, downstream workers generally do not correct the planning error themselves. The framework can verify executable code, but not subtle scientific validity; it may miss a wrong coordinate reference system or an unnecessary unit conversion. Retrieval quality remains metadata-dependent, and the archive paper notes that PANGAEA has sparse structured abstracts for only about 15% of datasets, making many datasets hard to discover (Pantiukhin et al., 24 Feb 2026).

Additional constraints include repository-specific evaluation, non-deterministic reasoning, interpretive hallucinations, cost, and security. The archive paper was tested on PANGAEA only, with transfer to EOSDIS, GBIF, and other repositories left unvalidated. The Python sandbox is deterministic, but the LLM-generated code is not; repeated runs can yield different code, ordering, or numerical choices. Sandbox isolation exists, but the paper states that there is no static code analysis or formal verification, and suggests AST checks, resource quotas, and import allowlists as future hardening measures (Pantiukhin et al., 24 Feb 2026).

The broader MAS Perspective frames these limitations as part of a larger research agenda rather than as isolated implementation issues. It emphasizes the lack of Earth-science benchmark suites, the difficulty of validating geoscientific visualization conventions, the need for specialized validators, and the challenge posed by domain-specific software cultures across ecology, biology, oceanography, climatology, and geology. It also proposes future directions such as re-exploring historical repositories, expedition planning, autonomous support for AUVs, self-organizing “wandering” agent systems, and more sophisticated centralized, decentralized, or hybrid MAS architectures (Pantiukhin et al., 7 Mar 2025).

A further misconception concerns nomenclature. PANGAEA-GPT is distinct from “Pangaea,” a separate general-purpose AI foundation model that encodes heterogeneous data as triplet sets and targets cross-modal transfer and myriad-task generalization rather than archive-native geoscientific workflow orchestration. The similarity in name reflects metaphor rather than identity of system design or application domain (Chang et al., 22 Sep 2025).

Taken together, the papers present PANGAEA-GPT as a repository-centered MAS for Earth and Environmental Science: not a replacement for scientific judgment, not merely a natural-language front end, and not a generic multimodal foundation model, but a workflow architecture that combines agentic search, metadata-driven routing, sandboxed deterministic execution, traceback-based self-correction, visual quality control, and narrative synthesis to make heterogeneous archives more discoverable, analyzable, and reusable (Pantiukhin et al., 24 Feb 2026, Pantiukhin et al., 7 Mar 2025).