What is agentic AI? From answering prompts to pursuing goals

What makes a system “agentic”

Agentic is a description of behaviour, not a single product feature. A handful of properties recur across systems people call agentic:

• Autonomy — the system decides its own next step within a goal, rather than a person specifying every action in advance.
• Tool use — it can call on external capabilities (a search function, a code interpreter, an API) rather than only generating text.
• Memory — it keeps track of what it has already tried and learned within a task, so it does not repeat itself or lose the thread.
• A multi-step loop — it plans, acts, observes the result, and re-plans, continuing until the goal is met, it hits a limit, or a person steps in.

Agentic AI versus an AI agent

The two terms are closely related and often used loosely, but it helps to keep them apart. Agentic AI is the broader paradigm — the behaviour of planning, acting and iterating with some autonomy. An AI agent is the concrete thing built on that paradigm: a specific system, wired up with particular tools and a particular goal, that behaves agentically. Put another way, “agentic” is the adjective and “agent” is the noun it describes; our companion explainer on AI agents covers how one is actually built, step by step. A single model call that just answers a question is not agentic. A system that plans several steps, uses a tool partway through, and reacts to what the tool returns, is.

Where agentic AI shows up today

The pattern has spread fastest wherever a task genuinely needs more than one step. Coding assistants that read a file, write a change, run the tests and fix what fails are agentic rather than a single code-completion call. Research tools that search several sources, compare what they find and assemble a summary behave agentically rather than answering from memory alone. Workflow automation uses the same idea to chain steps across systems — checking a record in one place, updating another, and notifying someone — that a person would otherwise do by hand, one click at a time.

Being honest about the limits

The autonomy that makes agentic systems useful is also where they go wrong. Because each step feeds the next, an early mistake tends to compound rather than stay contained — a wrong reading of a tool’s result three steps in can send the rest of the run in the wrong direction, and the longer a run continues, the more chances there are for that to happen. The underlying model can still generate a plausible-sounding but wrong step, and an agentic system acts on that step rather than merely stating it, which raises the stakes.

None of this argues against using agentic systems — it argues for guardrails: clear limits on what a system may do without a person checking in, logging that makes a run reviewable after the fact, and treating “can it act unsupervised” as a separate question from “does it usually get the right answer”.

How it relates to the rest of the AI stack

Agentic behaviour is a way of using a large language model, not a replacement for one — the model is still what reasons about each step, in the same way it reasons about a single prompt, just inside a loop that lets it act and re-plan. It is also a distinct idea from generative AI: generating a new piece of content is one possible action inside an agentic loop, but a system can be agentic (planning and using tools) without generating creative content at all, and it can generate content without being agentic.

Agentic AI in certification study

Agentic AI is an emerging area within AI-focused cloud certifications rather than a settled, heavily examined topic yet. The AWS Certified AI Practitioner (AIF-C01) expects a conceptual grasp of what agentic behaviour is and how it differs from a single model response, and Google Cloud’s Generative AI Leader certification covers the same shift towards autonomous, multi-step AI use at a similarly conceptual level. Neither exam expects you to have built one — they expect you to recognise the pattern and its trade-offs, which is exactly what this article sets out. Our /revision library carries the fuller syllabus for each exam, lesson by lesson.