Agentic Visual Reporting

A composable agentic system for automated visual data reporting, implementing a Human-AI Partnership model for transparent and steerable data analysis.

Overview

This prototype system demonstrates a multi-agent architecture that strategically externalizes logic from LLMs to deterministic modules, leveraging Draco for principled visualization design. The system delivers dual outputs:

• Interactive Observable reports with Mosaic for reader exploration
• Executable Marimo notebooks for deep, analyst-facing traceability

For complete details, see our whitepaper.

Architecture

The system consists of 11 specialized agents orchestrated in a multi-stage pipeline:

1. Data Understanding: Field Refiner → Dataset Describer → Field Expander → Dataset Profiler
2. Analysis: Insight Planner → Dataset Deriver → Dataset Publisher
3. Visualization: Dataset Visualizer (powered by Draco)
4. Reporting: Report Narrator → Dataset Reporter

Read more in our whitepaper.

Gallery

Explore example reports generated by the system in the gallery.

Getting Started

Ready to go deeper? Check out our Development Guide for complete setup instructions, API configuration, and usage examples.

Citation

If you use Agentic Visual Reporting in your work, please cite our workshop paper:

@misc{gyarmati2025ComposableAgenticReporting,
  author = {Gyarmati, Péter Ferenc and
            Moritz, Dominik and
            Möller, Torsten and
            Koesten, Laura},
  title  = {{A Composable Agentic System for Automated Visual Data Reporting}},
  year   = {2025},
  doi    = {10.48550/arXiv.2509.05721},
  url    = {https://visxgenai.github.io/subs-2025/3765/3765-doc.pdf},
  note   = {Workshop paper for the 1st Workshop on GenAI, Agents, and the Future
            of VIS at IEEE VIS 2025.}
}