What is agentic operations—and why do AI agents matter now?

Every six months, a new cohort of clinicians, public health practitioners and health-system executives arrives in our classrooms at Harvard Chan, and the questions they bring with them act as a reliable barometer of where artificial intelligence actually stands. Three years ago, most conversations centered on what generative AI was and whether a large language model could draft a passable clinical note or summarize a research paper. Eighteen months ago, attention shifted to prompt engineering and AI scribes. Today, leaders are grappling with a qualitatively different development: software that does not merely answer a question in a chat window, but takes a delegated goal, uses software tools across clinical and administrative systems, inspects what happens and decides what to do next.

That shift—from single-turn conversation to multi-step action—marks the arrival of agentic AI in everyday institutional life. Yet as the word “agent” has migrated from computer science papers into vendor pitch decks and hospital boardrooms, it has accumulated a thick fog of hype and imprecision. Products that run a rigid five-step script are marketed as autonomous agents, while genuinely capable systems fail in clinic because no one designed the surrounding workflow, permissions or human oversight. Understanding what agents actually are, why they have become practical right now and what it takes to run them dependably is the subject of a discipline we call agentic operations.

What an AI agent actually is

To answer this question, it helps to first separate predictive AI and generative AI from agentic systems. A predictive model takes structured inputs—vital signs, laboratory trends, prior admissions—and estimates a single probability, such as thirty-day readmission risk. A conversational chatbot takes a prompt and generates a response: a translated discharge instruction, a plain-language summary or a piece of computer code. Even when that response is remarkably fluent, the interaction begins and ends with generating text.

An agent, by contrast, is a system that pursues a (for now, delegated) goal across multiple steps, calls external tools, observes the results and adapts its subsequent actions. Imagine an outpatient specialty referral that has stalled. An AI model asked to draft a polite follow-up letter is generating text. A system that opens the referral record, notices that an outside pathology report is missing, checks whether the report arrived overnight in a document queue, routes a targeted request if it is still absent and logs its progress for a clinic coordinator is operating inside an action-and-feedback loop. As we note in our glossary guide to reading the names of models, harnesses, applications and providers, the underlying AI model may be identical in both cases. What turns a model into an agent is what it does and the software architecture wrapped around the model that enables agentic behavior.

In everyday organizational work, nobody is retraining a model’s internal parameters. Instead, the typically off-the-shelf model supplies interpretation, planning and proposed actions, while a surrounding AI harness supplies execution and control. The harness assembles the right information into the model’s context window, exposes approved capabilities through function calling and APIs, tracks current task state, carries relevant lessons across sessions through AI memory and enforces strict permissions and stopping rules. Autonomy is never binary: an agent can be given wide latitude to search internal guidance while holding zero authority to alter a clinical record or contact a patient without human approval.

Why agents have become relevant now

Why has this architecture moved from research laboratories into practical operations over the past eighteen months? Three developments converged at once. First, newer frontier models and reasoning models became substantially better at sustaining multi-step plans, inspecting error messages when a tool call fails and revising their approach rather than immediately losing the thread. Second, software engineering matured around the model: standardized protocols for connecting data sources, reusable agent skills, better context management and compaction and battle-tested harnesses—first proven in coding environments such as Claude Code, Codex, Goose and Antigravity—gave non-software industries a concrete blueprint for bounded multi-step work. Third, the cost of AI inference fell far enough that running ten or twenty reasoning and verification steps on a routine administrative case became economically plausible.

Those technical shifts arrived at a moment when health care badly needs operational relief, yet cannot afford careless experimentation. Clinicians are burning out on administrative friction; patients wait weeks while prior authorization packets, care transitions and diagnostic follow-ups stall between disconnected electronic health records; and health systems face acute staffing and financial pressures. Much of that friction lives in multi-step coordination work that is too unstructured for rigid twentieth-century health care automation, yet too high-volume for exhausted clinical staff to chase manually by spreadsheet and phone.

From impressive demonstrations to agentic operations

Here, however, lies the trap. Watching an agent complete a complex task in a five-minute demonstration is intoxicating, and it tempts organizations to ask the wrong question: What can this model do? In real clinical and public health institutions, a model capability is not a dependable service. Models remain stateless by default, uneven across closely related tasks and capable of AI hallucinations or confident compliance with flawed premises. Worse, if a team defines narrow volume metrics without careful boundaries, an agent can satisfy the dashboard while displacing rework onto nurses, leaking sensitive context or eroding AI safety.

Agentic operations is the discipline of organizing people, AI models and production software systems to achieve meaningful outcomes through bounded delegation, rigorous evaluation and continuous scientific learning. Rather than chasing autonomy as a trophy, it asks how work should be divided between conventional deterministic software, fixed AI workflows, bounded agents and accountable human professionals. It treats AI not as an oracle that replaces institutional judgment, nor as a passive widget bolted onto a broken process, but as a powerful instrument inside a governed system of work—where AI accountability never leaves human hands.

In the companion essays in this series—drawn from the teaching framework for our Harvard Chan executive course, Innovation with AI in Health Care: Agentic Operations—we walk step by step through how that discipline works in practice. We begin with purpose, scope and the three levels at which every operational problem must be diagnosed; examine what foundation models contribute and the four structural limitations every system must accommodate; look inside the production system’s software harness and operational infrastructure; explore the human layer of purpose, scientific inquiry and workforce stewardship; and close with the six-step operating discipline that turns everyday deployment into organizational learning.

Sources: Anthropic — Building Effective Agents (2024); Anthropic — Effective Harnesses for Long-Running Agents (2025).

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, including our executive course Innovation with AI in Health Care: Agentic Operations. 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