The convergence of HPC, AI, and edge instrumentation is enabling "Autonomous Science," a paradigm shift capable of compressing discovery cycles from years to months. However, shifting to agent-driven workflows introduces risks regarding non-determinism and "dataflow contamination," which threaten scientific reproducibility. In this presentation, we argue that robust provenance data management is the critical enabler for trustworthy autonomous systems.

We introduce a dual framework: "Provenance of Agents," which enforces accountability by systematically capturing agent decisions for root cause analysis, and "Provenance with Agents," which leverages Large Language Models (LLMs) as interfaces to democratize access to complex runtime data. Showcasing the Flowcept architecture and real-world applications in computational chemistry and adaptive additive manufacturing at Oak Ridge National Laboratory, we demonstrate how provenance safeguards the scientific method within autonomous loops. By ensuring transparency and enabling real-time human steering across the Edge-Cloud-HPC continuum , this approach removes manual bottlenecks, significantly reducing time-to-solution and accelerating the pace of trusted scientific discovery.