What is an AI agent?

The definition that actually works

An AI agent is a language model that takes actions in a loop to reach a goal it has been given.

That sentence has four things in it, and each one matters.

A language model. The brain is a large language model — Claude, GPT, Gemini, Llama, or any of the others. Without that brain, the system is not an agent. It is a script. The model is what decides what to do at each step.

Takes actions. An agent does not just answer your question. It does things in the world: searches the web, reads a document, sends an email, runs a calculation, books a meeting. Each of these is an action, and each action changes the state of something outside the agent.

In a loop. The agent does not act once and stop. It acts, observes what happened, decides what to do next, acts again, and so on. The loop continues until the goal is reached or until the agent gives up.

To reach a goal. Someone — usually you — has given the agent something to achieve. "Research three competitors and summarise their pricing pages" is a goal. "What is the capital of France?" is not — it is a question with one answer.

What an agent is not

It is useful to be specific about what does not count. A chatbot you have a conversation with is not an agent, even if it is powered by Claude or GPT. The chatbot answers each message in turn. It does not take actions in the world between messages, and it does not pursue a goal across messages. Each message is its own little task.

A script that uses an LLM to summarise a document is also not an agent. The LLM is being used as a clever function — text in, text out — but it is not deciding what to do next. The script decides; the LLM just helps.

A traditional automation tool — Zapier, IFTTT, or a cron job — is not an agent either. It follows rules a person wrote in advance. It does not decide.

Why this matters now and not five years ago

The agent idea is old. People have been writing about software agents since the 1990s. What changed is that language models became reliable enough to be the brain.

A few years ago, if you asked a model to plan a multi-step task, you got something incoherent. Now, the better models produce plans that often work on the first try. They can be given a tool — say, a search function — and they use it sensibly. They can recover when something goes wrong. They are not perfect, and we will spend a whole lesson on the ways they fail. But they are good enough that agents are a real engineering category, not a thought experiment.

This shift is the reason "agent" is everywhere in 2026. The brain finally works. The plumbing — function-calling, structured outputs, tool use, memory — has caught up. The remaining work is figuring out how to deploy these things responsibly.

A worked example: research vs answer

Imagine you ask Claude two questions, one after the other.

That was not an agent. Claude answered from what it already knew, in one turn, with one response. There was no loop, no action, no goal that needed pursuing.

Now imagine the second question.

If Claude tries to answer that directly from training data, the answer will probably be wrong. Restaurants close, change owners, and move. Some of them never existed. Claude does not actually know whether each restaurant is open on the Tuesday in question.

To answer that question well, the system needs to: search the web for current information about restaurants in that arrondissement, check the opening hours, check distances from a specific point, filter to vegetarian options, and write up the findings. That is several actions, with a loop, in pursuit of a goal. That is an agent.

The first task is something a model does on its own. The second is something an agent does, with the model at the centre.

The mental model to take with you

For the rest of this course, every time you see the word "agent", picture this: a language model sitting inside a loop, with access to one or more tools, working towards a goal you have given it. When you build your own agent in the lessons that follow, you will be building exactly this — and nothing more complicated.

Model answers

1. Web form — not an agent. No LLM, no loop, no goal pursuit. Just rules.
2. Customer support chatbot — not an agent in our sense. It is an LLM, but it answers one question at a time without taking actions in the world. Some chatbots do call tools (like looking up your booking), and at that point they become agents — but a pure Q&A chatbot is not.
3. Coding assistant — yes, agent. LLM brain, multiple actions (edits, runs tests), a loop (re-edit if tests fail), in pursuit of a goal (feature works).
4. Spam filter — not an agent. It is a classifier. No LLM, no loop, no goal beyond "is this spam".
5. Research tool — yes, agent. Classic case: LLM brain, multiple actions, a loop, goal-directed.
6. Lane-keeping — not an agent in our sense. It has a vision model and a loop, but no LLM, and the goal is reactive (stay in lane) not pursued through a plan.

If you disagreed with any of these, that is a good sign. The edges are genuinely fuzzy. Most working engineers would call (3) and (5) agents without hesitation, and would argue about (2).

Looking ahead

In Lesson 2 we will look closely at the loop itself — the think-act-observe cycle every agent runs. You will walk through a real agent run, step by step, and see exactly where the model decides what to do, what it does, and what it learns.