🤖 AI Agents, In Plain English (You Might Already Have One)

“Agent” is everywhere at the minute. Every AI newsletter, every second post on LinkedIn, and a load of tools that were a chatbot in January and have a shiny new badge on the pricing page now.

The explanations come in two flavours. One is a wall of words like orchestration and autonomous reasoning frameworks, written by someone who would like you to know they are clever. The other is a video of a robot butler booking a holiday in Lisbon. Either way you close the tab feeling like agents are a thing other people get to have, and you’ll catch up at some point.

So here’s the plain version. I teach this to business owners who have no interest in becoming engineers, and once you take the jargon off it there are about five moving parts.

Fair warning: by the end you might work out you’ve already built one.

What an agent is

A chatbot answers you. An agent goes and does the thing.

The definition, and there’s less to it than the hype suggests:

An agent is a model with tools, working in a loop, towards a goal you set.

The word doing the heavy lifting there is goal. You hand over the outcome, not the steps.

“Go through yesterday’s meeting notes, find anything I promised someone, and put it on my task list” is a goal. Notice everything you haven’t said. You haven’t said how many notes there are, or what to do about the call where you promised nothing, or which of the four Sarahs is the right Sarah. It works that out on the way.

What one is made of

Five parts, and you already understand four of them.

The model. The brain. Claude, or whichever one you’re using.

The instructions. The goal, plus your rules. What it’s for, and what it must never do.

The tools. Its hands. Your calendar, your inbox, a document, a spreadsheet, a search box. When the model reaches out and uses one, the technical term is tool use, which is a rare case of AI jargon meaning exactly what it says on the tin.

The loop. The engine. More on this in a second, because it’s the whole ballgame.

The harness. The bit nobody explains, and the reason people stay confused. It’s the scaffolding wrapped around the model: it runs the loop, holds the memory, and hands over the tools the model is allowed to touch. Claude, the app you type into, is a harness. So is anything else with a chat box and a connectors menu.

Think of a kitchen. The model is the cook, the tools are the knives and the oven, and the harness is the kitchen itself: the worktops, where things are kept, and the rule about which cupboard stays shut.

Why the loop is the bit that matters

Plan. Do the next step. Look at what came back. Go again. Stop when the goal is met.

That’s the loop, and it’s the difference between an agent and every automation you’ve set up that broke the second something looked slightly different.

A script does the same steps in the same order no matter what. Point it at a document laid out differently and it falls over, or worse, it doesn’t fall over and quietly does the wrong thing for a fortnight.

A loop reads the result of each step and picks the next one based on what it found. So when the notes are a mess, or a name is spelled two ways, or one meeting has nothing in it worth logging, it can adapt instead of breaking.

It can also catch itself. Not always, and nowhere near reliably, but an agent will sometimes look at what it produced, clock that it isn’t right, and have another go. A chatbot never gets that chance, because it has already handed you the answer and gone quiet.

Knowing when to stop is part of it too. A goal has a finish line: task list updated, inbox triaged, report written. It works until it gets there, then it stops.

The reframe: you might already have one

This is the part nobody says plainly.

A scheduled task is a whole little agent. It has a goal, it has tools, it makes its own decisions about the steps, and it stops when it’s done. If you’ve set one up to read your inbox at 7am and write you a summary, congratulations, you built an agent. Nobody sent you a certificate.

A skill is a component. A ready-made ability sitting on a shelf waiting to be picked up. Reusable, handy, not an agent on its own.

A connector is its hands. Access to your Drive, your calendar, your Notion. Also not an agent on its own.

Which is why a lot of what gets sold as “agentic” turns out, when you open it up, to be a chatbot with a connector bolted onto it. Worth knowing before somebody invoices you for it.

The free plan is fine for trying Claude and getting a feel for how it thinks. Connectors, Skills and scheduled tasks are what the paid plan (Pro, around £20 a month) is for, so the thing that runs while you’re asleep is the paid bit.

The honest bit

What they’re brilliant at: boring, repeatable, multi-step jobs with an obvious finish line. Filing, chasing, summarising, tidying, moving things from where they landed to where they’re supposed to live. The work you’d hand to a very fast assistant who has no memory of last Tuesday and no feelings about being asked twice.

Where to keep a hand on the wheel:

They are wrong with total confidence. A made-up answer turns up in the same calm tone as a correct one, and there’s no wobble in the voice to tip you off. Calling it an agent doesn’t fix that.

They inherit your mess. Woolly instructions produce woolly work. Point one at a chaotic folder and it will use the chaotic folder, cheerfully, at speed.

Anything with real consequences stays supervised. Money moving, emails going out to clients, anything you can’t undo. Draft, don’t send, is a rule worth keeping long after you feel ready to drop it.

Where this is heading

The useful direction is duller than the demos and better for it. Less of the invisible admin needing a person: the filing, the chasing, the writing-up, the “who did I say I’d get back to on Thursday”. Work that never shows up on an invoice and eats your week anyway.

That’s the part quietly moving off people’s plates. It gives you hours back, which is less exciting than a robot butler and worth considerably more.

So, today

Nothing to build. Go and look at what you’ve already got: any scheduled task, any connector, any skill sitting in your account. The parts are all there and now you can name them.

Nick this one:

I'm working out which bits of my current AI setup count as an agent, and which are components of one.

Here's what I've got: [list it: connectors you've switched on, any skills, any scheduled tasks, and roughly what each one does]

For each one, tell me:
1. Whole agent, or a component? One line on why.
2. Which part it is: the model, the instructions, the tools, the loop, or the harness.
3. If it's a component, what it would be a component of.

Then pick the one job I've described that has the clearest finish line and the smallest consequences if it goes wrong.

I'm not asking how to build anything yet, I want to see what I already have. Plain English, under 250 words.

If the vocabulary around all this is what’s been putting you off, the rest of it is in Every AI Word That’s Been Confusing You, Explained.

That’s what an agent is. Building one around your own work, your clients, your prices and your actual week is the next thing entirely, and it’s what The Claude Kickstart sets up: Claude taught your real job, five jobs off your plate, and a playbook for your trade. £47, one weekend.