Agentic Large-Language-Model Systems in Medicine: A Systematic Review and Taxonomy

Abstract

In less than three years, large language models (LLMs) have advanced from passive responders to agentic systems capable of planning, acting, and collaborating with other agents. In medicine, these capabilities open the door to systems that can assist clinicians, coordinate care, and adapt to complex, real-world workflows. But the high-stakes nature of healthcare demands a careful examination of their potential: understanding not only the range of applications but also the challenges of data privacy, ethical responsibility, and safe deployment. This survey systematically analyzes over 140 studies from 2022 to 2025 on LLM-based medical agents, offering three key contributions. First, we introduce a unified taxonomy spanning application domains, autonomy levels, and integration of tools and knowledge, effectively categorizing use cases across various biomedical areas. Second, we explore how state-of-the-art agents integrate biomedical expertise, couple LLM reasoning with external resources like clinical databases, electronic health records (EHRs), APIs, and implement human-in-the-loop mechanisms to mitigate hallucinations and biases. Third, we synthesize crossdomain insights from fields such as education, robotics, and automated scientific discovery, highlighting transferable design principles for improved reliability and interpretability. Our analysis identifies persistent challenges, including hallucinations, biases, data privacy risks, and regulatory complexities, and reviews promising solutions to address them. Finally, we propose open research directions toward creating trustworthy, regulation-compliant agents that augment rather than replace clinical expertise, providing researchers, practitioners, and policymakers a comprehensive roadmap for advancing agentic AI in healthcare. A comprehensive list of agentic AI models studied in this work is available at here