Human-Centered Agentic AI for Biomedical Research: A Framework for Scientific Agency, Evidence, and Governance

Agentic artificial intelligence systems are extending biomedical AI beyond prediction and text generation toward multistep workflows that can plan, retrieve evidence, use tools, execute code, evaluate intermediate results, and, in some settings, connect computational reasoning with laboratory research. These capabilities may reduce coordination costs and expand access to specialized methods, but they also create new risks: errors can propagate across workflows, persuasive outputs may conceal weak evidence, and tool access can turn incorrect inferences into consequential actions. This perspective presents a human-centered framework for evaluating and governing agentic AI in biomedical research. It distinguishes technical capability, scientific validity, and institutional authority; proposes a multidimensional taxonomy of research autonomy; and introduces an autonomy–consequence approach for calibrating oversight. It also develops a lifecycle with four human decision gates—problem formulation, planning, result validation, and release—a multidomain assurance model, and a five-phase institutional roadmap spanning mapping, sandboxing, validation, governance, and staged scaling or retirement. The central argument is that human oversight should be designed as a scientific function rather than added as a final approval step. Responsible adoption requires proportional permissions, independent verification, reproducible provenance, contestability, rollback mechanisms, and continuing evaluation of the complete human–agent system. This article is a conceptual perspective and narrative synthesis; no new data were generated or analyzed.