Medical AI Agents: A Comprehensive Survey of Architectures, Cognitive Modules, and Clinical Workflows

Abstract

Agent-driven multi-agent systems, built upon large language models (LLMs), are emerging as a promising paradigm to address the challenges of complex data and accelerated decision-making in medicine. However, despite rapid progress, research in this area remains fragmented across narrow scenarios, making it difficult to systematically assess and adopt these technologies. To address this gap and synthesize the fragmented landscape, this survey introduced a task-oriented taxonomy that organizes medical AI agents into four core application landscapes: Clinical natural language processing, Medical Computer Vision, Graph Learning, and Omics and Biology. Drawing on recent studies published mainly between 2023 and 2025, this review characterized the roles that agents serve for clinicians and patients and discussed how existing systems integrate reasoning, memory, and tool use into coherent pipelines aligned with concrete clinical workflows. The availability of code, models, and agent implementations is also documented to inform practical reuse. Furthermore, the review synthesized cross-cutting challenges reported in the literature, including data quality, clinical grounding, safety, transparency, and deployment. Finally, it outlined future directions for building robust, controllable, and clinically actionable medical AI agents.

Ähnliche Arbeiten