What Is an AI Agent, Really?

Directions versus the parcel

There are two favours you might ask a colleague on a Tuesday morning.

The first is where is the post office. They tell you. Left at the lights, past the bakery, grey building with the unfortunate blue door. Genuinely useful. You still have to put your coat on and walk there yourself.

The second is could you go and collect my parcel. They stand up, leave, queue behind a man posting a canoe paddle, sign something, and come back holding a box. Somewhere in the middle they made a dozen small decisions you never heard about and never needed to.

That is the entire difference between a chatbot and an agent. One gives you directions. The other comes back with the parcel.

Almost everything else you have been told about agents in the past two years is decoration on top of that one idea.

What changes when it does instead of says

A chatbot has exactly one move: produce text and hand it to you. Ask it to book a table and you will get a lovely paragraph about booking a table. Nothing has been booked. Nothing outside the conversation has moved a millimetre.

An agent gets the same request plus a set of things it is allowed to actually use: a search box, a calendar, a document, an inbox. It picks one, uses it, looks at what came back, and decides what to do next. It keeps going until the job is finished or it gets stuck. (A sibling course, Understanding Loops in AI, takes that repeat-until-done cycle apart properly, so we will leave it alone here.)

The practical consequence is worth sitting with for a second. A chatbot that misunderstands you wastes ninety seconds of your time. An agent that misunderstands you emails the wrong document to the wrong client. Identical misunderstanding, very different Tuesday.

Three things an agent has that a chatbot does not

A goal, not a question. You do not ask an agent what it thinks. You tell it what should be true when it stops.

Tools it can reach. Without tools an agent is just a chatbot with ambition. The next lesson is entirely about this.

Permission to continue. A chatbot answers once and waits. An agent takes a step, sees the result, and takes another one without asking you first. That permission is the thing that makes it useful and the thing that makes it risky, in exactly equal measure.

When it finished, had anything changed outside the chat window?
No  = chatbot
Yes = agent


Did it take more than one step on its own?
No  = chatbot
Yes = agent


Could it have got the job wrong in a way you would need to undo?
No  = chatbot
Yes = agent, and it needs boundaries

The word is doing a great deal of work

Be warned: agent is now a marketing word. It gets stapled onto anything with a text box. In practice there is a spectrum, and most products sit somewhere in the middle of it rather than at either end.

ANSWERS ONLY       Writes you a draft. You copy, paste, send.
SUGGESTS ACTIONS   Prepares the email and shows it. You press send.
ACTS WITH A CHECK  Does the work, pauses at the risky step, asks you.
ACTS ALONE         Does the whole job. You see the result afterwards.

Ask it like a chatbot and you get chatbot results

The most common mistake with an agent is talking to it the way you talk to a chat window. Questions get answers. Agents need a destination.

Where this leaves you

You do not need to know how any of this is built. You need to know one thing when someone offers you an agent: what can it touch, and what happens if it gets that wrong. That single question will carry you through the rest of this course.

Pick a job with these three properties:

It is boring and you do it repeatedly
Every step of it is easy to check afterwards
If it goes wrong, you can undo it in under five minutes
Good candidates: tidying a list, summarising a folder of documents,
drafting (not sending) routine replies, collecting links on a topic.
Bad first candidates: anything involving payments, customers,
public posts, or deleting things.