What Is an AI Agent? A Plain English Explanation

The short answer

A chatbot answers you. An agent does things for you. Give it a goal rather than a question and it can take several steps on its own, using tools such as web search, your files or your email, checking its own progress as it goes and having another attempt when something does not work.

The interesting part and the risky part are the same part: it acts without asking you between each step. That is what makes an agent worth having, and it is also why the only question that really matters is what the thing is allowed to touch.

What is the actual difference between a chatbot and an agent?

Think about two favours you might ask a colleague.

The first is asking for directions to the post office. They tell you: left at the lights, past the bakery, grey building with the regrettable blue door. Useful. You still have to put your coat on and walk there.

The second is asking them to go and collect your parcel. They stand up, leave, queue behind a man posting a canoe paddle, sign something, and come back holding a box. Along the way they made a dozen small decisions you never heard about and mostly did not need to.

Directions versus the parcel. That is the whole distinction, and almost everything written about agents in the past two years is decoration on top of it.

The practical version is less romantic. A chatbot that misunderstands you wastes ninety seconds. An agent that misunderstands you sends the wrong document to the wrong client. Same misunderstanding, very different afternoon.

What does an agent have that a chatbot does not?

Three things, and you can check for all of them before you hand anything over.

A goal instead of a question. You do not ask an agent what it thinks. You tell it what should be true when it stops. "Summarise this" is a question. "Read these twelve files and produce one page listing every deadline, sorted by date" is a destination.

Tools it can reach. A tool is one specific thing it is allowed to do: search the web, open a document, look at a calendar, send a message. Without any tools connected, an agent is a chatbot in a costume.

A loop. It takes a step, looks at the result, decides whether that got it closer, and takes another one. Nothing about this is mysterious, but it is where most of the surprises come from, which is why we gave it a whole free course on how AI loops work.

How do you tell a safe agent from a risky one?

Ignore the word agent on the homepage entirely and look at the permissions screen. Then sort everything it wants into two columns: reading and writing.

An agent that reads and proposes is a research assistant. It looks things up, opens files you pointed it at, gathers what it found and hands you a draft. When it gets something wrong, you notice while reading, and the cost of the mistake is that you read a paragraph you did not need. That is a recoverable category.

An agent that writes is something else entirely. Sending, publishing, booking, editing, paying and deleting all leave a mark on a world other people can see. When it gets one of those wrong, you are not correcting a draft, you are ringing someone up on Monday morning to explain.

This is the single most useful distinction for a normal person, and it is worth more than any amount of technical detail. Not how clever is it. Not which model does it use. Just: after this thing has finished, what is different, and can I put it back?

What are agents genuinely good at today?

Four categories, all of them unglamorous, all of them real.

Research with a defined finish line. Go and find twenty things, compare them against criteria you specified, come back with a table. Tedious for a person, well suited to something that does not get bored.

Sorting and routing. Reading a pile of incoming things and deciding which bucket each belongs in. Tickets, enquiries, applications, receipts. The judgement is shallow and repeated, which is exactly the shape that works.

Drafting from several sources. Pulling a summary out of a folder of documents, three spreadsheets and a chain of emails. The value is not the writing, it is not having to hold six things in your head at once.

Repetitive multi step admin. The jobs that take four minutes but involve six windows, and you do them eleven times a week.

Notice what all four have in common: the output is something you read before anything happens. That is not a coincidence.

Where do agents actually fail?

Here is the failure that should worry you, and it is not the dramatic one.

A small error at step two becomes a confident conclusion at step nine. The agent misreads a column heading, or picks up last year's version of a file, or interprets a vague instruction slightly sideways. Then it builds on that, competently, for seven more steps. What arrives on your desk is well organised, fluently written, correctly formatted and wrong in a way that is now buried three layers down.

Nobody watched the middle. That is the actual problem. You saw the brief you gave and the answer that came back, and the interesting bit happened in between where there was no one in the room. Our course on checking AI work is largely about learning to spot this before it reaches a client.

Agents also fail in a quieter way: they do not know when to stop. Given a goal it cannot reach, an agent will often keep trying, keep spending, and eventually produce something that technically answers the question while missing the point entirely.

What does a trustworthy agent run look like?

You will not watch every step, and you should not have to. What you can insist on is an honest report at the end.

WHAT I DID        The steps taken and anything changed, in plain language
WHAT I COULD NOT  The file that was locked, the page that would not load,
the thing you asked for that does not exist
WHAT I ASSUMED    Every gap it filled in on your behalf without asking

The third line is the one people skip and the one that matters. Assumptions are where a helpful agent and a useless one part company. A tool that says "I could not find Q3 figures so I used Q2" has given you something you can check. A tool that quietly uses Q2 and never mentions it has given you a problem with a delay fuse in it.

Do you actually need one?

Probably not yet, and that is a perfectly respectable answer.

Most of what people want from agents, they can get from a good prompt and thirty seconds of their own attention. A well briefed chatbot beats a badly scoped agent almost every time, because the badly scoped agent takes nine steps in a direction you did not intend and presents the result as finished work.

Where agents earn their place is when the same shaped job comes round every week, every step of it is easy to verify, and getting it wrong costs you five minutes rather than a client. Start there. Start in read only. Widen the boundary once you have seen what it does with a narrow one.

If you want the longer version, with tool permissions, briefing, failure modes and how much rope to give it, our free Introduction to AI Agents course walks through it properly. If you would rather see agents doing real work first, AI That Does the Work is the practical companion. And if you are not sure where to begin at all, start here.