What an AI Agent Actually Is (and When You Actually Need One)

Tomasz Dylik
6 days ago
5 min read

The word agent is on every second page you read about AI now, and almost nobody stops to say what it means.

If you have been nodding along and quietly hoping that somebody would eventually explain it, you are not the only one. I put off learning this properly for months, because every explanation I opened was written for people who already build software.

So here is the version I wish I had read first. It ends with one small thing you can do today.

The word covers four different things.

That is the whole reason it feels slippery. Four quite different kinds of software are being sold to you under one word, and they are not equally useful, equally expensive, or equally risky.

It helps to put them in order, from the simplest to the most independent.

Level one is rules. You write down what should happen, and it happens. If an email arrives from your bank, mark it important. If somebody fills in the form, add them to the list. It never learns and it never surprises you. That is a feature, not a flaw. Most of the automation quietly running your life sits here.

Level two is an assistant. This is ChatGPT, Claude and the others you already know. It understands what you meant rather than only what you typed, and it writes something genuinely good. Then it stops and waits. Every next step needs you to ask for it. Ask well and you get a lot. Ask with no direction and you get the most average answer in the room.

Level three is an agent. Here is the real difference: you stop describing the steps and start describing the destination. You give it a goal and the rules it has to respect. It works out the steps itself, does them one at a time, looks at what came back, and adjusts.

Level four is a team of agents. Several of them, each with one job, and one that hands out the work. A researcher, a writer, an editor. This is real, it works, and it is almost certainly not where you should start.

What makes the third one different.

An assistant answers. An agent finishes.

To finish something without you standing over it, a system needs four things, and if one of them is missing you are back at level two wearing a costume.

It has to remember, not only the last few messages but the things it was told earlier and the documents you gave it.

It has to plan, meaning it can take one large instruction and turn it into an order of small ones.

It has to be able to reach things outside the conversation. Your calendar, your inbox, a spreadsheet, a website. An assistant can write you an email. An agent can send it.

And it has to keep going, checking after every step whether it is closer to the goal, and trying something else when it is not.

That last one is the one people underestimate. It is also where the cost lives, because every loop is another step you pay for, and another place something can go sideways.

The same job, done two ways.

Say you want a week of email dealt with.

With an assistant. You open a message, paste it in, ask for a reply, read it, copy it back, send it. Then the next one. You are doing eleven small jobs and the assistant is doing one of them eleven times. It is faster than writing from scratch. It is still your whole morning.

With an agent. You write down, once, what counts as a simple question, what counts as one that needs you, and how you like to sound. The system reads the inbox, sorts the messages by what they are actually asking for, drafts answers to the simple ones, and puts the three awkward ones in front of you with a note about why. You spend twenty minutes on the three that were always going to need judgement.

Same task. The shape of your work has changed completely.

The free AI Agents pack. If you would rather look before you build, I put the practical parts in one place: over two thousand ready-made n8n workflows you can open and read, a workbook on the twenty building blocks that cover most of them, the Agent-Plus Prompt Builder for writing agent instructions, and the checklists I run through before letting anything work on its own. Download the AI Agents pack →

When you actually need one.

Now the honest part, because this is usually where an article starts selling and I would rather not.

Most people reading this do not need an agent. They need a better prompt and somewhere sensible to keep it.

An agent earns its keep when three things are true at the same time. The task repeats, often enough that you are tired of it. It takes more than one step, so there is something worth handing over. And it happens when you are not there, or at least when you would rather not be.

Two out of three is a maybe. One out of three is a prompt you should save, not a system you should build.

There is a fourth test, quieter than the others. Can you write the rules down? Not roughly. Actually write them, in sentences, so that a stranger could follow them without needing to ask you anything. If you cannot, the agent will not manage it either, and it will not tell you. It will simply do something you did not intend, politely.

What an agent will not do for you.

It will not use judgement you never wrote down. Everything you know but never said stays outside the system.

It will not notice that it is wrong. It has no feeling of doubt. If it invents an order number, it hands you that number with exactly the same confidence as a true one, unless you have deliberately given it a source it has to check against.

It will not stay finished. Passwords expire, websites move things around, a connected tool renames a field. Anything that runs without you needs looking at every so often. That is not a defect. That is what owning a small machine is like.

And it will not turn a slow, unclear process into a fast, clear one. It will make it faster and just as unclear.

One small thing to do before the next post.

Write down three tools you used today. Any three. The alarm on your phone, your bank app, ChatGPT, the sat nav, the rule that files your newsletters.

Next to each one, write a number from one to four.

It takes two minutes, and it does something that no amount of reading does. The word agent stops being fog and becomes a position on a short list. Most of your three will be ones and twos. That is what I found when I did it, and it is the reason I stopped feeling behind.

Go deeper on each part.

If you want the whole thing in one place.

If you would rather have the complete system than one piece at a time, that is what I wrote AI Agents Bible for. Five books in one: what agents are, how to build them on Custom GPTs, Make.com and n8n, thirty worked examples you can copy, and the parts most guides leave out, about cost, reliability and knowing what not to let run on its own.