A Composable Agentic System for Automated Visual Data Reporting

Creating insightful visual data reports involves multiple decision points and iterative refinement. While Large Language Models (LLMs) show promise in automating this process, monolithic agentic approaches often prove brittle in practice. We present a prototype system that addresses this challenge through a human-AI partnership model. Our multi-agent architecture strategically moves logic away from language models into deterministic modules, specifically leveraging the Draco visualization design system. The system produces dual outputs: an interactive Observable report with Mosaic for exploration, and executable Marimo notebooks for analyst-facing traceability. This design creates an auditable and steerable system that balances automation with human oversight, demonstrating how compositional architectures can enhance both reliability and transparency in automated data reporting.