Your agent writes the code. You build what it can’t get past.
A course in building real software with AI, for people who have never written code. An agent is an AI that can change files and run programs on a computer. It works fast and it’s sometimes wrong with total confidence, so you learn how to check it.
Start at chapter 1: The machineAgent
Code an AI really wrote, on its first try (Claude Haiku 5.5). A recording, not a mock-up.
Production
Stands in for real users: hidden checks the agent never saw say whether it works.
Each moving dot is an attempt by the AI, heading for users. With no gates, all of it gets through, broken or not. Press a gate to add a check (on a desktop you can also drag the cable onto one), then watch what it stops, and what good code it blocks by mistake. These gates are what this course teaches you to build. Every attempt is real, from our experiment, and every verdict is what that tool really said.
That line of gates is part of a harness: everything around an AI model that decides what it sees and what happens to its work. Learning to build one is the difference between someone who ships what an agent wrote, and someone who can tell you why it’s right.
No experience is assumed. There are 18 chapters, drawn on this site as wires, and 526 words, most with a working instrument you can press and break. It starts at what a program is and ends with you defending your work to someone paid to doubt you.
Every word comes with a sentence you can type to your agent, because you can only ask for what you can name.
The route
New to all of this? Start at chapter 1 and go in order, because each one relies on those before it. Each chapter ends with a stage of the build, one real app you make with an agent as you go, and that repo is what you show at the end. We estimate 27 to 44 hours of your own time for the build, on top of reading, and say on that page how we got the number. If you already know some of it, the chapters are labelled, so start where you are.
Before the agent
You can't supervise what you can't read.
The agent
What it is, what it sees, and how to tell it what you want.
The harness
The part that makes the output trustworthy.
Everything it can get wrong
The ways software fails that tests rarely catch by themselves.
What you’ll be able to say, and show
The big line is what you could say to an employer or a client. The smaller text is what you could show them, and you’ll learn every word in it on the way. Anyone can say the first, and interviewers now ask for the second. It only works if you really do the work on your own machine. The drills at the end of each chapter practise one thing, and the build puts it all in one repo, with what to point at for each line below.
- My agents write the code.
- A project with a short instructions file for the agent, where you can trace each line back to a mistake an agent once made. Learn this
- But nothing goes in on their word.
- Automatic checks on a server (called CI) that every change must pass, covering its logic, its style, leaked passwords and the outside code it uses. And a rule that stops the agent calling a job finished while any of them fails. Learn this
- I know the tests really test.
- A test is a small program that checks your code. A mutation score breaks the code on purpose and counts how many breaks the tests notice. You show that score, the gaps you closed, and checks the agent never saw. Learn this
- I can read what it wrote.
- Any file, explained with the agent switched off, and a bug you found by hand with a debugger, a tool that pauses a program so you can look inside it. Learn this
- I know what it costs and what can go wrong.
- A threat model (who might attack this, and how), an agent that cannot reach the live system, and a figure for what a change costs in tokens, the units AI use is billed in. Learn this
- And when it breaks, it is mine.
- A written postmortem, an honest account of what failed and why, and the check you added so that the same failure gets caught next time. Learn this
The workshop
- The buildOne real app, a bill splitter called Kitty, built with an agent from an empty folder to a postmortem. One stage per chapter, with the exact words to paste and what went wrong when we ran it first.
- Review gymReal code changes that passed every automatic check, each with a real bug hiding in it. Train your eye until you find them, with every line explained in plain words if you need it.
- Interview roomThe questions someone paid to doubt you will ask, from gentle to brutal. What they are testing, what a weak answer sounds like, and what to show instead.
- Your repoEach gate you pass adds real, working files. By the end you can download the harness you earned and put it on your own project.
- The experimentWe ran a real agent hundreds of times to test this course’s own claims. Some of what we expected turned out to be wrong, and the page says so.
- Who said whatWhere each idea came from: Hashimoto, Lopopolo, Böckeler, Tan, Martin, Willison, Beck, Huntley and more. And the popular misquotes that would embarrass you.
- ReceiptsThis site was built by agents, under the harness it teaches. Here is every check that ran on it, and what each one caught.