Agentic AI is a class of system in which a language model is given a goal and the authority to decide how to reach it — choosing actions, calling tools, observing the results, and adjusting its next step. Rather than producing one output from one prompt, it operates over many steps, exercising a degree of autonomy that distinguishes it from a single model call.
Why agentic AI matters
A single model call maps one input to one output. That is enough to draft an email or answer a question, but many real tasks are not one-shot: booking travel, triaging a support queue, or resolving an incident require gathering information, taking actions, checking what happened, and deciding what to do next. The exact sequence depends on facts that are not known when the task begins.
Agentic AI addresses this by letting the model drive a multi-step process. Instead of a developer hard-coding every branch in advance, the system reasons about the current state and selects the next action. This makes it possible to automate work that is too variable for fixed scripts — but it also introduces a control problem. An autonomous, multi-step process that calls external systems can fail partway through, repeat side effects, or stall waiting on a person, and those failure modes are the central engineering challenge of running it in production.
How it works
An agentic system is usually built from a small set of cooperating parts:
- A model that reasons over the goal and the history so far and proposes the next action.
- Tools the model can call — search, code execution, database queries, external APIs — that let it gather information and affect the world.
- A loop that runs the chosen action, feeds the result back to the model, and repeats until the goal is met or a stop condition is reached.
- Memory and state that carry context across steps so the system knows what it has already done.
The loop is what makes the behavior agentic: each iteration is conditioned on the outcome of the last, so the path through the task is decided as it unfolds rather than fixed in advance.
Agentic AI vs. generative AI
Generative AI and agentic AI are often used interchangeably, but they describe different behaviors. Generative AI is concerned with producing an output — text, an image, code — from a prompt in a single forward pass. Agentic AI uses a generative model as one component inside a control loop that takes actions and reacts to their results. Put differently, generation answers a question, while an agentic system pursues a goal, and pursuing a goal is what creates the need for durability, observability, and guardrails.
In practice
Running an agentic system reliably means treating the loop as a process that must survive crashes, redeploys, and pauses — not as a function call that either returns or fails. A durable, observable runtime persists each step so the run can resume after a failure, and records what the model decided and why. The same building blocks reappear at every scale: an AI agent is the unit, an agentic workflow composes its steps, and orchestration coordinates several agents working together. For a deeper rationale, see why durable agents.