Abstract
Agentic artificial intelligence (AI) is beginning to change scientific data work from a sequence of manually connected tasks into a set of tool-mediated, partially autonomous workflows. That shift creates a practical question: can agents accelerate analysis without weakening the findability, accessibility, interoperability, reusability (FAIR), and reproducibility on which scientific claims depend? We systematically reviewed 218 peer-reviewed papers, preprints, and community standards published from 2016 to September 2026. The evidence was organized into five linked domains: data engineering and AutoML; semantic data models and knowledge graphs; multimodal analytics and retrieval-augmented generation; provenance and reproducibility; and human oversight, trust, and governance. Across these domains, the literature shows clear progress in execution-grounded data preparation, workflow orchestration, semantic retrieval, and multimodal evidence synthesis. It also shows a consistent boundary: systems can generate plausible analyses faster than they can establish that those analyses are complete, reproducible, and scientifically warranted. The most persistent weaknesses are incomplete lineage, unstable dependencies, weak cross-modal semantics, inconsistent evaluation, and poorly calibrated human reliance. We therefore frame trustworthy agentic science as a data-to-decision problem rather than a model-selection problem. Agents should produce claim-level provenance, machine-actionable semantic metadata, executable environments, uncertainty-aware outputs, and explicit escalation points for human review. The review concludes with a staged research agenda and an integrated framework that links FAIR practice, reproducibility engineering, and human governance across the scientific lifecycle.