This paper presents a novel approach for analyzing massive amounts of provenance data generated from scientific data processing workflows spanning edge, cloud, and HPC environments. To overcome the limitations of existing approaches, which rely on custom scripts, structured queries, and static dashboards for data interaction, we propose a runtime data analysis methodology, reference architecture, and open-source implementation leveraging an interactive Large Language Model (LLM) agent. A lightweight, metadata-driven design transforms natural language into structured provenance queries. Evaluations on real-world chemistry workflows using various LLMs, including LLaMA, GPT, Gemini, and Claude, demonstrate that the modular design, prompt tuning, and Retrieval-Augmented Generation (RAG) enable accurate and insightful responses beyond recorded provenance.
