Intelligent Agents in Biomedicine: Architectures, Applications, and Implications for Neuropsychiatric Imaging

Abstract

The complexity of neuropsychiatric disorders demands a paradigm shift from static, single-task AI models to dynamic frameworks capable of integrating multimodal data across the full continuum of diagnosis and treatment. This review introduces the agentic AI paradigm, which is characterized by environmental perception, memory, planning, tool use, and adaptive interaction, as a framework uniquely suited to these challenges. We trace the evolution of intelligent agents from symbolic AI and cybernetics to architectures driven by large language models (LLMs), and synthesize a general four-module architecture comprising perception, memory, planning, and action. Current biomedical agent applications are organized into a hierarchical classification spanning task-oriented, workflow-oriented, and research-exploratory paradigms, covering advances in imaging analysis, clinical decision support, research coordination, and drug discovery. Focusing specifically on neuropsychiatric imaging, we provide a narrative review of current AI and LLM-based methods, revealing that existing approaches overwhelmingly remain at the task-oriented level, with few published systems having yet integrated the full suite of agent capabilities within a closed-loop architecture. We analyze domain-specific challenges in data quality and heterogeneity, interpretability and theoretical validity, clinical safety and accountability, and ethical governance. Finally, we outline future directions including brain-inspired multi-agent architectures and the concept of specialized probe agents designed not merely as diagnostic tools but as computational instruments for testing mechanistic hypotheses about brain–behavior relationships. This review aims to provide researchers with a foundational understanding of agentic AI and a roadmap for its principled application in neuropsychiatric imaging.