Agentic AI in medicine: architectures, applications, evaluation, and challenges for clinical translation

Medical artificial intelligence is evolving from single-task predictive models—such as those designed solely for image classification or risk scoring—toward agentic systems. Unlike traditional models that map inputs directly to outputs, agentic AI architectures incorporate planning, memory, tool use, and iterative feedback. These systems are designed to handle complex, multistep clinical workflows by coordinating specialized agents, retrieving external medical knowledge, and interacting with clinical databases or diagnostic tools.

This scoping review analyzed 557 studies to map the current state of agentic AI in healthcare. The research identifies a broad range of applications, including medical question answering, automated image interpretation, electronic health record (EHR) analysis, and clinical trial prediction. These systems often utilize frameworks like ReAct or multi-agent collaboration to emulate multidisciplinary consultation, allowing for more nuanced decision-making than standalone foundation models.

Despite the rapid development of these architectures, the review highlights a significant gap between technical capability and clinical readiness. Most existing studies rely on public benchmarks, simulated settings, and retrospective data. Critical dimensions for clinical adoption—such as process reliability, evidence traceability, error recovery, and safety—are inconsistently evaluated. The authors argue that for agentic AI to move into real-world clinical workflows, the field must shift toward reproducible evaluation protocols, auditable oversight mechanisms, and prospective validation in actual healthcare environments.