What an AI agent actually is

“Agent” is doing a lot of work in a sentence these days, most of it vague. It gets used for a chatbot, a script with an API key, and a research demo that books flights, as if they were the same kind of thing. They are not, but they do share a shape, and once you can see the shape the marketing stops being confusing.

An agent is three parts. A model that produces text. A set of tools it is allowed to call. And a loop that runs the model, reads what it asked for, does it, and feeds the result back in — until the task is done or a limit is hit. That is the whole idea. Everything else is an implementation detail on top of those three.

The model on its own can only write. Ask it for the weather and it will produce a sentence that looks like a weather report, invented. It has no way to check. The model is the part that decides what to do; it is not the part that does anything.

This is the distinction most “AI” products blur. A model behind a text box is a model behind a text box. It becomes an agent only when you give it a way to act on the world and a reason to loop.

A tool is a function with a description the model can read: search the web, run this query, send this email, read this file. The model does not run the tool — it emits a request to run it, in a structured form, and your code decides whether and how to honour that request.

That gap is the whole safety story. The model proposes; your code disposes. A well-built agent treats every consequential action — spending money, sending a message, deleting anything — as a request that passes through a gate you control, not a command that executes because the model felt like it.

The loop is the part that earns the name. Run the model. It asks for a tool. Run the tool. Give it the result. Run the model again, now with more information than it had a second ago. Repeat until it says it is finished, or until a step budget stops it.

This is why an agent can do things a single prompt cannot: it gets to react. It can search, read what it found, notice it was wrong, and search again — the same way a person works a problem they do not fully understand at the start. It is also why agents fail in the ways they do. A bad loop wanders, repeats itself, or confidently acts on a misread. The intelligence is real; the judgement is only as good as the tools, the gates, and the limits you gave it.

When someone shows you an “AI agent”, you now have three questions that cut through the demo. Which model? What tools can it actually call? And what happens in the loop when it is wrong — who catches it before it acts? The answers tell you far more than the word ever will.