In recent years, with the rapid advancement of artificial intelligence, we are at a pivotal transition from large language models (LLMs) to multimodal large language models (MLLMs), where AI can deeply understand text, images, audio, and environmental information. However, in healthcare settings, what is required goes beyond question answering—it involves decision-making and execution. This lecture explores how medical AI is evolving from response-based systems into medical agents with capabilities in planning, reasoning, and autonomous action, and even toward embodied AI that can interact with the physical world.

The talk will systematically examine the three generations of medical AI development:

First generation (Discriminative AI): classification models focused on image recognition and disease prediction.

Second generation (Generative AI): conversational assistants for doctor–patient interaction and clinical documentation.

Third generation (Agentic AI): intelligent agents capable of autonomous tool use and complex clinical pathway planning.

Using virtual multidisciplinary consultation as an example, we will demonstrate how medical multi-agent systems (MAS) overcome the limitations of single models by simulating collaborative workflows of professional healthcare teams. Finally, the lecture will address key challenges in deploying agent-based architectures in digital medicine, including reasoning reliability, semantic alignment across multimodal data, and accountability in medical decision-making, while outlining a new paradigm of healthcare driven by agentic AI.