GeoFlow: Agentic Workflow Automation for Geospatial Tasks

Amulya Bhattaram, Justin Chung, Stanley Chung, Ranit Gupta, Janani Ramamoorthy, Kartikeya Gullapalli, Diana Marculescu, Dimitrios Stamoulis

TL;DR

GeoFlow is a new framework that automates the creation of multi-agent workflows specifically for complex geospatial tasks. By embedding explicit, detailed tool-calling objectives into task graphs, it achieves a 6.8% improvement in task success and up to a fourfold reduction in token costs compared to current state-of-the-art orchestration methods.

What Problem Does This Paper Address?

Geospatial Information Systems (GIS) are notoriously complex, involving hundreds of specialized API tools, diverse satellite data products, and strict parameters like coordinates and time ranges. While Large Language Model (LLM) agents show promise in navigating these systems, manually designing their workflows (who talks to whom and in what order) is a bottleneck that doesn't scale.

Existing "automated" workflow generators often rely on implicit reasoning. For example, a meta-agent might tell a sub-agent to "run a detector," but in a GIS environment with dozens of different detectors (SAR vs. optical, ship vs. building), the sub-agent often lacks the context to choose the right tool or parameters. This leads to "hallucinated" tool calls or execution errors. GeoFlow aims to bridge this gap by making the orchestration instructions explicit and context-aware.

Methods

GeoFlow builds upon the Activity-on-Vertex (AOV) graph paradigm, where nodes represent subtasks and edges represent dependencies. The process follows a two-stage pipeline:

  1. Workflow Generation: A "meta-agent" LLM receives a user request and generates a JSON-based AOV graph. Unlike previous methods, GeoFlow prompts the meta-agent to populate each vertex with a highly specific "agentic scope and objective." Instead of a vague instruction, the objective includes the exact API name and required parameters (e.g., "Run the Swin-L EO detector for dates 2024-XX over AOI Z").
  2. Workflow Execution: The system performs a topological sort on the graph to determine the execution order. Sub-agents are then invoked via standard function-calling, using the detailed objectives as their primary context. This ensures that even if the global chat history is long or messy, the agent knows exactly what its specific job is.

Datasets

The researchers evaluated GeoFlow using the GeoLLM-Engine benchmark, which consists of 20 realistic Earth Observation tasks. These tasks involve object detection and land cover classification across millions of satellite images.

To test the robustness of the method, they ran evaluations across several major LLM families, including:

  • OpenAI: o3 and o4-mini.
  • Qwen: Qwen3 14B and 8B.
  • Meta: Llama 3.3 70B.
  • Mistral: Mistral 3.1 24B.

Key Insights and Ideas

  • Explicit Context Prevents Drift: The "aha moment" here is that sub-agents fail primarily due to ambiguity in their return conditions or scope. By "baking" the specific API requirements into the task definition itself, GeoFlow prevents agents from "over-reasoning" or attempting actions outside their assigned subtask.
  • Efficiency over Consensus: Traditional multi-agent designs (like AutoGen's Group Chat) require significant "chatter" between agents to reach a consensus on the next step. GeoFlow’s graph-based approach eliminates this overhead, leading to massive token savings.
  • Open-Source Parity: One surprising finding was that the Qwen3 14B model achieved the highest success rate (77.3%), outperforming even some proprietary frontier models. This suggests that specialized geospatial automation is highly achievable with efficient, non-proprietary models.

Key Results

  • Higher Success: GeoFlow achieved an average success rate of 74.2%, a 6.8% improvement over the previous state-of-the-art "Flow" method (67.4%).
  • Cost Reduction: Compared to AutoGen’s Group Chat, GeoFlow reduced token usage by over 4x (6.2k tokens vs. 27.1k tokens) while maintaining higher accuracy.
  • High Correctness: The "Flow Score" (a metric measuring how well the generated graph matches a human-crafted ground truth) was consistently above 95% across all tested models.

Limitations and Future Work

The authors acknowledge that the current GeoLLM-Engine benchmark tasks are largely linear in their logic. Real-world geospatial analysis is often iterative, requiring "human-in-the-loop" adjustments.

To address this, the authors propose integrating GeoFlow with no-code interfaces (like AutoGen Studio). In this vision, the AI generates the initial AOV graph, and a human GIS analyst can visually inspect, edit, and refine the workflow before execution. Future research will focus on these more complex, multi-round interactive tasks where the environment (maps and databases) changes dynamically.

Summary updated

Abstract

We present GeoFlow, a method that automatically generates agentic workflows for geospatial tasks. Unlike prior work that focuses on reasoning decomposition and leaves API selection implicit, our method provides each agent with detailed tool-calling objectives to guide geospatial API invocation at runtime. GeoFlow increases agentic success by 6.8% and reduces token usage by up to fourfold across major LLM families compared to state-of-the-art approaches.
BibTeX
@inproceedings{bhattaram2025geoflow,
  author = {Amulya Bhattaram and Justin Chung and Stanley Chung and Ranit Gupta and Janani Ramamoorthy and Kartikeya Gullapalli and Diana Marculescu and Dimitrios Stamoulis},
  title = {{GeoFlow: Agentic Workflow Automation for Geospatial Tasks}},
  booktitle = {SIGSPATIAL/GIS},
  year = {2025},
  url = {https://openreview.net/forum?id=Ct3FfnTx9I},
}

Loading related content...

Figure