How do I create an agent in ChatGPT?

Checked on 22 October 2026 against each maker’s own help pages.

How can I create my own AI agent?

If you open ChatGPT and explain the same things every time (who you are, what the job is, how you like it written), you are ready for an agent. That is all most people mean by the word: an assistant you brief once, which keeps the brief.

In practice it is a set of standing instructions, a name and, if you want, a few documents to work from. You write the instructions, try them on a real task, correct them and save. Every conversation with it then starts already briefed.

No code is involved. The skill is briefing a new member of staff well, and most of us are worse at that than we think.

How do I get to the agent builder — and why can’t I see it?

There are two ways to brief it once. Set either up on a computer, in the web version or the desktop app. The simpler is a project: a folder for related chats, with its own standing instructions and files that every chat inside it starts from. The fuller is what the maker now calls a plugin: you describe the repeated job in an ordinary conversation, add your instructions and reference files, test it and save it. The older builder for custom GPTs, the one most guides describe (mine included), is being retired, and the maker is moving those to plugins. The menus move often, so I am not naming the buttons; the maker’s own help pages have today’s route.

If you can’t see it, it is usually one of four things:

• You are on the phone app. Build on a computer; use it anywhere afterwards.
• Your plan does not include building. See the next question.
• You are on a work account and your administrator has switched it off. Ask them. Don’t route round them.
• You are looking for the other thing. “Agent” is also used for the mode in which ChatGPT carries a larger task through in steps by itself, using tools such as a browser. When I checked, the maker called that mode Work, with its own switch. It is not the builder.

Is the ChatGPT agent free?

Not all of it, and I won’t print a price here, because it changes. My 2023 guide said building one was for paying users. Today, by the maker’s own help pages, the cloud browser the step-by-step mode uses is not on the free plan or the lowest-priced paid one, and which plans can build what keeps changing. What your plan allows is on the maker’s own plans page. Read it there, on the day.

The more useful question is whether your organisation already pays for a business account. Building inside it is usually the better route: your administrator can see it, and on business plans your data is normally kept out of training by default. Read that line in your own contract before you build anything that touches work.

What can a ChatGPT agent actually do?

One repeated job, done well, if you brief it well. The kinds that earn their keep:

• drafting in your house style, so the first version already sounds like you;
• answering from a set of documents you give it (a handbook, a policy, a price list) instead of from everything it has ever read;
• a first pass on something you will then check: a summary, a reply, the bones of a report.

The step-by-step mode goes further: it can open pages, click and fill things in. That is where behaviour matters most. An AI given a task and permission to act will look for a way through whatever is in front of it. It doesn’t know what a door is (Door Finding & Key Making). Give it the smallest permission that does the job.

Can I build one in Copilot or Claude instead?

Yes. Both let you build an assistant with standing instructions and your own documents, under different names and different menus. In Copilot it is called an agent, built from the Copilot chat in your work account on a computer, if your licence and your administrator allow it. Copilot is the likely route if your organisation already runs on that maker’s office software, because the agent can work from files you already keep there. In Claude it is a project, with its own instructions and documents, and projects are open on the free plan too, up to a limit. I run my own work on Claude.

The choice matters less than people think. The brief, the checking, and the ways AI goes wrong at the edge of what it knows are the same in all three. Choose the one your organisation already pays for and can see into, and spend your effort on the brief.

What should I tell my agent first?

Its purpose, in one sentence. That was the best tip in my 2023 guide and it still is. Start with “Your job as my assistant is to…” and finish the sentence plainly. Don’t pour in everything you think you want. Build it up.

Then add, in this order:

1. Who it works for, and who will read what it writes.
2. What it must never do: send anything, invent a source, guess a figure.
3. What to do when it doesn’t know: say “I don’t know”. It rarely does this unless told, and not always then.
4. How it should talk: plain, UK spelling, no flattery.
5. What it may work from, and what it may not.

Try it on something you already know the answer to. When it gets it wrong, correct the instructions, not the conversation. A correction made in a chat lasts for that chat; one made in the instructions lasts. The same habit stops you getting stuck in a huge thread you daren’t leave.

How do I stop my agent making things up?

You don’t. You reduce it, and you catch it.

An agent is the same machine underneath, with a brief on top. When it reaches the edge of what it knows, it does what the chat window does: it carries on, fluently, in the same confident tone as the answers that were right. No brief removes that.

What reduces it: documents to work from, an instruction to answer only from them and to say when something isn’t there, and a narrow job. What catches it: a person who reads the output against the source before anything goes anywhere.

That is the gap between what the machine does and what you do, and holding it is your job, not the agent’s. Four checks you can run today: Why AI makes things up.

When do we need someone to design it for us?

When it stops being one person’s assistant. The signs:

• more than one person relies on it;
• it touches client work, or anything that leaves the building;
• it does several jobs that hand work to each other;
• nobody can say who signs off what it produces.

At that point the question is not which button to press. It is what the room around the agent looks like: what it is for, what it may touch, what it must never do, and which named person reads what comes out. That is what I call agentic design. It is done inside your own AI account, with the maker you already use. No engineering, no hosting: nothing of mine is installed, and nothing of yours comes to me. How agentic design works.

Paul Roebuck · Psychotherapist, author and AI coach. I have written 14 million words with AI (counted September 2026), and I train leaders on how it behaves.

Want your team to brief and check their agents well, or want one designed properly? Book a 20-minute chat.