How AI agents actually work, explained without the jargon

Every week there is another headline about AI agents. Almost none of them explain what an agent actually is. So here is the plain version, the way I would explain it to a mate down the pub.

An agent is four things

Strip away the marketing and every AI agent is the same four parts:

A model. The clever bit. Something like Claude or GPT that can read, write, reason, and plan.

Instructions. A job description. Who it is, what it does, what it must never do, how you like things done. This is written in plain English and it matters more than the model.

Tools. Access to things: your email, your calendar, your files, the web, your task list. A model without tools can only talk. A model with tools can work.

A place to work. Somewhere it runs continuously, not a browser tab you open and close. A server, a spare Mac Mini, a platform like Buzz.

That is it. Model, instructions, tools, somewhere to run. Everything else is packaging.

The useful analogy

ChatGPT is a clever person you can phone. Knows a lot, gives good answers, but cannot touch anything and forgets you the moment you hang up.

An agent is an employee with keys to the office. It has a job description, it can open the filing cabinet, and it is still working when you have gone home.

Both are useful. But only one of them clears your inbox before you wake up.

What agents are genuinely good at

Reading and triaging email. Scheduling. Drafting replies for you to approve. Watching tools like Slack and surfacing what matters. Research. First drafts of almost anything. Repetitive admin that follows a pattern.

The pattern is: high volume, low judgement. Anything where the task is boring but the standard matters.

What they are bad at

Judgement calls. Anything involving money moving without approval. Reading the room. Knowing when a client needs a human. New situations with no pattern to follow.

A good setup respects this. My agents draft, I approve. They surface, I decide. Get that division wrong in either direction and the whole thing fails: too little access and it is a toy, too much and it is a liability.

What setting one up actually involves

Three things, in order of how long they take:

Deciding what it should do. An afternoon of honest thought about where your time goes.

The technical setup. A day, if you know what you are doing. Accounts, access, security lockdown, testing.

The instructions. Ongoing. The job description gets better every week as you correct it and it learns how you work. This is the part nobody tells you about, and the part that separates a demo from a team member.

That is the honest picture. Not magic, not hype. A well set up agent is the cheapest productive hire you will ever make. A badly set up one is a weekend project you abandon. The difference is entirely in the setup.