Agentic AI
Most software runs the same fixed steps every time. Agentic AI is what you get when some of those steps aren't fixed — the model decides what happens next, instead of the code deciding for it.
It's not a type of system. It's a measure, a scale from none of that to a lot of it. A system that's scripted except for one small decision barely counts. A full AI agent, which decides its whole sequence of actions on its own, sits at the far end of the same scale.
Three points on the scale, least to most agentic
Respond: the model answers once and stops. Ask it why a test is failing, and it explains: "the function looks like it returns the wrong value, try changing X." Then it stops — a person does the actual fix.
Make one decision: the model gets to choose one thing — say, whether it needs to search documentation before answering — but the application still controls everything else.
Work through a task: given the goal "fix the failing test," the model reads the error, edits the code, and reruns the test itself. If it still fails, it inspects further and tries again; if it passes, it reports the result. This is the AI agent end of the scale — decide, act, check the result, decide again, until the goal is reached.
More agentic isn't automatically better
A predictable task — extract the fields from an invoice and save them — has known steps, so a fixed workflow handles it better: cheaper to run, easier to test, and it can't make a judgment call it wasn't supposed to make. An unpredictable task — find out why checkout conversion dropped — doesn't have a knowable path in advance; that's where letting the model decide the next step actually helps, because no one could have scripted that exact sequence ahead of time. Being agentic doesn't mean unrestricted, either — a system can decide its own steps and still need a person's approval before sending money or deploying code.
Agentic AI vs. AI agent
An AI agent is one specific point on this scale — the far end, fully autonomous. "Agentic AI" is the broader idea of the scale itself: every AI agent is agentic, but not everything agentic is a full agent. It's also not the same as generative AI — generating text or images means producing something, agentic means deciding and acting toward a goal, and a system can do either one without the other.
Practice interview questions on agentic AI →
In this guide
FAQ
Can a system become more agentic over time without anyone redesigning it?
Not on its own — moving along the scale means changing what the system is allowed to decide, which is a deliberate design change (handing the model a choice it didn't have before, or removing a check that used to block it). A system doesn't drift toward more autonomy by accident; someone has to widen what it's permitted to decide.
Can two systems use the same underlying model but sit at different points on the scale?
Yes — the model itself doesn't determine where a system sits. Two products built on the same model can differ entirely in how much of the decision-making the surrounding application hands over versus keeps fixed in code. The scale describes the system's design, not the model's capability.