JMIR Publications

ABSTRACT

Deep Research Agents are autonomous LLM-based systems capable of iterative web search, retrieval, and synthesis. They are increasingly positioned as the next major leap in medical AI. In this Viewpoint, we argue that while these agents mark progress in information access and workflow automation, they represent an incremental evolution rather than a paradigm shift. We review and summarise current applications of Deep Research Agents in medical scenarios, including literature review generation, clinical evidence synthesis, guideline comparison, and patient education. Across these early use cases, the tools demonstrate the ability to rapidly gather and structure up-to-date information, often producing outputs that appear comprehensive and well-referenced, offering tangible benefits to clinicians and researchers. Yet these strengths coexist with unresolved and clinically significant limitations. Citation fidelity remains inconsistent across models, with subtle misinterpretations or unreliable references still common. Their retrieval processes and evidence-ranking mechanisms remain opaque, raising concerns about reproducibility and hidden biases. Moreover, overreliance on AI-generated syntheses risks eroding clinicians’ critical appraisal skills and may introduce automation bias at a time when medicine increasingly requires deeper scrutiny of information sources. Safety constraints are also less predictable within multi-step research pipelines, increasing the risk of harmful or inappropriate outputs. We contend that Deep Research Agents should be embraced as powerful assistants rather than pseudo-experts. Their value lies in accelerating information gathering, not replacing rigorous human judgement. Realising their potential will require transparent retrieval architectures, robust benchmarking, and explicit educational integration to preserve clinicians’ evaluative reasoning. Used judiciously, these systems could enrich medical research and practice; used uncritically, they risk amplifying errors at scale.