Software people keep asking what “agentic AI” means, and most answers either dodge the question or drown it in jargon. Here is a short one, with a worked example, and an honest section at the end about when you should not reach for this at all.
Start with a moment rather than a definition.
An email lands in a shared inbox: “Invoice 4471 does not match what we ordered.”
A chatbot can write a good reply to that email. It can be polite, it can ask for the purchase order number, it can sound like it is helping. What it cannot do is open the purchase order, open the invoice, put the two side by side, find the line that differs, and tell you that you were billed for 200 units of something you ordered 20 of.
Closing that gap is what people mean by agentic AI, most of the time, when they are not simply relabelling a product.
The definition
An agentic AI system is software that uses a language model to decide what to do next, takes that action against the real world, looks at what came back, and decides what to do after that. The decisions are not hardcoded. The model is in the loop.
That is it. Everything else is plumbing.
Back to the invoice. An agent handed that email decides to look up the purchase order, and your code runs the lookup. It reads what came back and decides it now needs the invoice, so your code fetches that too. It compares the two, finds the quantity mismatch, and decides the next step: either correct the record or write a reply that names the exact line and the exact numbers. Nobody wrote that sequence in advance. The model chose each step after seeing the result of the last one.
What it is not
A plain chatbot does not qualify. It receives a message and produces a reply. There is no follow-up action, no observation of effects, no second decision. The model produces words and stops.
A generative AI app that lets the user click “use this tool” is closer, but still not agentic. The human is the one closing the loop, choosing the tool, judging the output, deciding whether to continue. Take the human out and the system stalls.
An agent closes its own loop. It picks the tool. It reads the result. It decides what to do next, including stopping.
Why the word is suddenly everywhere
Two things happened at once, and it helps to separate them.
The first is real. Language models got reliably good at one narrow, unglamorous skill: reading a list of available tools and emitting a correctly formatted request to call one of them. Tool calling went from a research trick to a standard feature of the major model APIs. Once a model can pick a tool and get the arguments right most of the time, building the loop around it becomes worth the effort. Before that, the loop broke too often to ship.
The second is marketing. A great deal of software that was already shipping got relabelled. Workflow automation became “agentic workflows.” A chatbot with a search box became an “agent.” This is why the term feels slippery: a fair amount of what carries the label does not close its own loop, and some of it never will.
Both things are true at the same time. The capability shift is genuine and the vocabulary around it is inflated. Working out which one you are looking at is the whole skill, and it is the subject of what counts as agentic AI, and what does not.
What changes when the loop closes
When the agent closes its own loop, the system can do multi-step work without a person at every junction. That is the promise, and on the invoice it is concrete: three lookups, one comparison and a drafted reply, with a human reading the result instead of assembling it.
It is also where most of the engineering problems live. The agent will sometimes pick the wrong tool. It will sometimes mis-read a result and act confidently on the wrong picture. It will occasionally call the same tool over and over, waiting for a different answer. None of that is exotic; it is the normal behaviour of a system that decides for itself, and building a real agent is mostly the work of bounding it so the upside is reliable enough to ship.
That work is the difference between a demo and something you can put in front of customers, and it is why most agentic AI demos fail in production.
A useful first test
If you are looking at a system and unsure whether it is “really” agentic, ask one question. After the model produces output, does that output get fed back into the model along with new information from the environment, and does the model then decide what to do? If yes, it is an agent. If no, it is a generator with extra steps.
When agentic AI is the wrong answer
This is the part that usually gets left out, and it matters more than the definition.
When the steps are the same every time. If the task never varies, a script is cheaper, faster, and it does not surprise you. Putting a model in the loop buys you the ability to handle variation. If there is no variation, you are paying for judgment you do not need.
When the answer has to be identical every time. Pricing, tax, regulatory disclosures, anything an auditor will read. A model that decides for itself is the wrong shape for work where “usually correct” is a failure. Use ordinary code and let the agent do the messy parts around it.
When being wrong is expensive and nobody can check the work. An agent that drafts a reply for a person to read is a different risk from an agent that issues a refund on its own. The second one needs limits, a log you can read afterwards, and a clear answer to who gets paged when it goes wrong.
When you cannot say what a good answer looks like. This is the one that sinks projects quietly. If you cannot describe the difference between a good outcome and a bad one, you cannot test it, and an agent you cannot test is an agent you cannot safely run. Deciding which work is actually a fit is its own exercise, covered in what to automate first.
None of these are reasons to avoid agentic AI. They are the reasons a sensible first project is smaller than the one people usually propose.
Where to go next
If this is the first definition that has landed for you, the beginner book walks through the same idea one chapter at a time, building a single project (an internal IT helpdesk assistant) from a chatbot into a real agent.
For the mechanics of the loop itself, read the four steps of the agent loop. To see it in working Python in about 150 lines, read Build an AI Agent From Scratch in Python.
Two companion posts pick up where this one stops. This page answers what agentic AI is. What counts as agentic AI, and what does not answers the harder question you get in a real meeting, which is whether the specific thing on the screen qualifies, with an autonomy ladder and a checklist you can run. The agentic AI glossary defines the forty surrounding terms, so you can look up one word without reading an essay.
Frequently asked
Quick answers
- Is a chatbot an agentic AI?
- No. A chatbot takes a message in and produces a reply. It does not pick a tool, take an action against the real world, or read the result and decide what to do next. Without that loop closing without a human in the middle, it is not agentic.
- How is agentic AI different from generative AI?
- Generative AI generates content. You give it an input, it gives you an output, it stops. Agentic AI uses a generative model as the decision engine inside a loop that takes real actions, observes results, and decides again. Every agentic system uses generative AI; not every generative AI system is agentic.
- Is an LLM the same as an agent?
- No. An LLM is the language model itself. An agent is a system built around the LLM that gives it tools, lets it pick which one to call, runs the tool, feeds the result back, and lets the LLM decide what to do next. The LLM is one component of an agent, not the agent itself.
- Can an agent work without internet access?
- Yes, if its tools do not need the internet. An agent that searches a local knowledge base, runs commands on the local file system, and uses a locally-hosted model does not need network access. Most useful production agents do reach the internet because that is where the tools they need live.
- Are agents going to replace software engineers?
- No, but they are changing what engineers do. Engineers spend less time typing each line of code and more time defining intent, reviewing what the agent produced, and bounding what it is allowed to do. The job is moving up a level of abstraction, not disappearing.