What is agentic AI?
The AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
Agentic AI describes systems in which AI helps choose actions across a sequence of work. The system can use what it discovers to decide what to do next, rather than only produce a single response or follow a wholly predetermined sequence.
A language model that drafts a reply is generating text. An application that checks a record, notices missing information, searches another source and revises its plan is doing something more: using a model within an action-and-feedback loop. The model may be the same in both applications. The difference lies partly in the software and permissions around it.
Generative and agentic AI therefore overlap. An agent may generate a search request, a plan or a draft message. The categories are not competing model types, and a long conversation is not automatically an agentic process.
A practical agent needs a goal, relevant context, available tools and a stopping rule. Context is the information supplied for the current decision. Tools are functions that obtain information or perform actions. State records what has already happened, while memory may bring relevant information into later tasks. Not every agent needs an elaborate version of each component.
There is no need to make every step flexible. Ordinary software can check required fields, calculate amounts and enforce permissions. A model can handle the parts that require interpretation or choosing between possible next steps. The simplest design that performs the task dependably is often the best starting point.
In clinical operations, consider a referral that has stalled because a report is missing. An agent could identify the gap, check whether the report has arrived elsewhere and prepare a request for an authorized coordinator. A complete evaluation asks whether the referral progressed correctly, not merely whether the agent produced convincing messages. It also examines duplicate requests, incorrect recipients and unresolved exceptions. Agentic AI is useful when adapting the route helps complete the work; its value should be demonstrated in the resulting service.
Sources: Anthropic — Building Effective Agents (2024); Yao and colleagues — ReAct: Synergizing Reasoning and Acting in Language Models (2023)
To learn more, take a look at the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy