Dr. Anthony Chang

“Just as Classical music elevated Baroque complexity into human expression, agentic AI elevates deep learning’s power into dynamic, goal-directed intelligence.”

Anthony Chang, violinist and AI advocate + ChatGPT 4o

One of the hottest topics right now in artificial intelligence is agentic artificial intelligence, or agentic AI. Little is known (or understood) by clinicians and non-technologists, however, about the complicated and sophisticated AI technology and architecture that will need to be assembled for agentic AI to be executed anywhere near its (over)exuberant marketing.

Interestingly, agentic AI does share a philosophical and technological lineage with cognitive computing: both of these AI methodologies focus on thinking and reasoning, but there are also important different nuances. This concept of a “human-like” strategy enabled by a collection of tools is reminiscent of cognitive computing promulgated by IBM last decade (but hopefully with much better longer term outcome). While cognitive computing (exemplified by IBM Watson Oncology Advisor) possessed a myriad of technologies like machine learning, natural language understanding and processing, information retrieval, automated reasoning, rule-based systems, and knowledge graphs; agentic AI has an equally large collection of AI tools such as foundation models along with memory, planning, and tool use systems (detailed below). Perhaps agentic AI can be thought of as the next evolutionary step of cognitive computing, with the additional processes of planning, acting, and learning beyond thinking.  

First, agentic AI is not just another AI tool like classifiers in machine learning or large language models with transformers; agentic AI is more of an ensemble of systems to achieve its main objective: to act with autonomy towards goals. This collection of AI systems or modules for agentic AI comprises of:

1) Foundation models- large language models (LLMs), vision models, and multimodal models will give the AI agents general understanding (e.g. GPT-4o, DeepSeek, Claude 3, or Gemini)

2) Planning modules- specialized subsystems that can plan multi-step actions and use methods such as chain/tree of thought reasoning and search, task decomposition, or even traditional symbolic AI

3) Memory systems- short-term (“working memory”) and also longer-term memory (“episodic memory”) to remember past actions and learn over time with vector databases, knowledge graphs, or memory embedded inside models

4) Tool-use modules- systems that enable the AI to extend itself by calling external systems, APIs, and databases or performing other functions (such as run queries, manipulate files, browse the Internet, control apps, etc)

5) Reflexion systems- these tools enable AI agents to set their own goals, critique their own work, and adjust their plans for situations by using tree of thought, self-critique, or reflexion with an agent “loop”.

Second, the AI technologies needed to execute the above main objective and collection of tasks are some of the most advanced AI technologies to date (an agentic AI “all-star” technological tool team). The portfolio of technological tools necessary to execute the agentic AI objective includes:

1) Transformers- the foundational technology of agentic AI are these large parameter models with self-attention

2) Reinforcement learning- agents to improve at achieving goals can be accomplished with reinforcement learning of various types, such as reinforcement learning with human feedback (RLHF) and its newer congener reinforcement learning with AI feedback (RLAIF)

3) Fine-tuned models- AI agents can be built on top of models that will be fine-tuned for desired agent behavior (such as “do this”, “plan that”, etc)

4) Retrieval-augmented generation (RAG)- this tool is used for extracting information and knowledge dynamically during task execution

5) Toolformer and tool-augmented learning models- these models are trained to recognize when exactly the models will need to call a tool

6) Multi-agent systems- multiple AI agents will need to interact, collaborate, and discuss how to execute the tasks in a cohesive manner with either communication protocols amongst agents or orchestration layers to choose the best AI agent.

Third, agentic AI is different than the other AI tools in that agentic AI will follow “execution loops”, and not at all like “one-and-done” models that responds merely once to a single prompt. The execution loops are usually the following or some variant of this scheme: think > plan > act > observe (outcome) > reflect > adjust > repeat. This is, of course, very similar to how humans achieve goals.

Agentic AI Examples 

Some of the early available agentic AI tools include:

1) LangChain- this tool chains LLM actions together to perform complex tasks but it is not considered agentic AI; it is a framework that helps to build agentic AI

2) Autogen (Microsoft)- this is an open-source framework that allows multi-agent collaborative systems to solve problems and is much closer to being agentic AI than LangChain

3) AutoGPT- allows users to automate multistep projects and complex workflows with AI agents based on LLMs, and is considered a tool with agentic behavior in LLMs.

More recent agentic tools include SuperAGI, CrewAI, and DeepSeek Agents. Examples of the very few agentic (or near-agentic) AI deployed in healthcare include:

1) Hippocratic AI (early stage for low-risk healthcare tasks at low level of autonomy) for patient interaction

2) Aidoc for radiology workflow agents and it is moving towards agentic AI (currently considered to be semi-agentic AI)

3) Qventus specializes in real-time hospital operations management and is considered an example of agentic AI

4) Insilico Medicine and Benevolent AI are relatively strong agentic AI systems in the domain of scientific discovery

The future

Additional future areas for agentic AI to develop include real long-term memory, meta-reasoning and meta-cognition, multimodal agency, self-improvement, and ideally embedded ethical, legal, and regulatory layers with human input and oversight. Agentic AI is in its very nascent stages in clinical medicine and healthcare but emerging in clinical documentation, operations optimization, patient communication, research support, and clinical workflow automation and wherever else dynamic planning and tool use are of paramount importance.

Overall, agentic AI will be welcomed to mitigate the very large burden in the imbroglio of healthcare around the world. We need to be careful to heed the lessons from the earlier great promises (but never fulfilled) of AI technological advances, especially cognitive computing a decade ago. It is easy to be focused (and distracted) on the technological aspects of agentic AI rather than the implementation and application of the AI technology to solve problems in healthcare. Finally, it is also very easy for non-clinicians to (vastly) underestimate the level of difficulty of the complex and unpredictable nature of the problems in healthcare.  

Agentic AI and its challenges will be among the popular topics covered at AIMed25, the world’s longest-running event dedicated to artificial intelligence in medicine and healthcare. Since 2013, AIMed has brought together over 1,000 clinicians across 50+ subspecialties, along with healthcare leaders, data scientists, trainees, investors, and innovators from around the globe.

This year’s 3-day meeting, taking place November 10–12, 2025 at the Manchester Grand Hyatt in San Diego, features dedicated tracks on AI in pediatrics and neonatology, AI in health professional education, and AI and mental health. Attendees can also expect hot-topic breakfast workshops, specialty breakout sessions, an abstract competition with scholar awards, the popular American Board of AI in Medicine (ABAIM) course, and—for the first time—a Chief AI Officer agenda.

Join us as we explore the future of healthcare, including the role of AGI, generative AI, large language models, intelligent XR, and much more. We look forward to seeing you there!