What a Coding Assistant Is

Waymark wants a small web page. A guest types their booking reference into a box, and the page tells them whether the trip is confirmed, waitlisted or cancelled. Tessa asks Milo's agency what that would cost, and Milo says two days — then adds, as though it explained itself, "the assistant writes most of it now."

She has been hearing that sentence for two years without ever pinning it down. This page pins it down. A coding assistant is the model from this book — the same model, on the same provider's computers, reached by the same request Chapter 7 took apart — pointed at code instead of prose. Code is the text a program is written in: instructions in a language the computer follows exactly, which is what Tessa was looking at when Milo's twenty lines were read out loud in Chapter 8.

The surprise is not that a text machine can produce code. It is how well the fit works, and why.

Why Code Suits the Model

Two facts do the work here, and both of them come back on every remaining page of the chapter.

The first is that code has stronger patterns than almost any other kind of text. Chapter 1's mechanism — given some text, produce text that fits — has an easier target when the text is code, because code is written to rigid conventions by people who copy each other on purpose. There is a normal way to write a loop, a normal way to name a thing, a normal shape for a page that takes a reference number and looks it up. A booking-status page has been written some enormous number of times, in public, and the model has read a great deal of it. Waymark's version is not a new idea. It is a familiar idea with Waymark's details in it.

The second fact is the one that makes code genuinely different from everything else in this book: code can be run. When the model drafts a brochure paragraph, nothing outside Tessa's own judgement can tell her whether it is any good. When the model writes a page, the computer will attempt to do exactly what the code says, and the attempt either works or falls over with an error message. That is an objective check, arriving in seconds, of a kind the model's other outputs have never had.

Hold that second fact carefully, because it is about to be overstated everywhere else you hear about this. "It runs" is not the same as "it is correct". A program can run perfectly and do the wrong thing — count cancelled bookings as confirmed, forever, without a single error message. Running proves the instructions are well formed. Whether they were the right instructions is a separate question, and the next three pages are about what happens when nobody asks it.

The Ladder: Three Rungs

"Coding assistant" covers three quite different things, and telling them apart is most of what Tessa needs in a vendor meeting.

The lowest rung is autocomplete. As the programmer types, the assistant proposes the rest of the line, or the next few lines, in grey. One keystroke accepts it; carrying on typing ignores it. The human is writing the code and the machine is finishing sentences, exactly the way a phone keyboard finishes yours.

The middle rung is chat over code. Paste in some code, describe the problem, get an explanation or a revision back. This one Tessa can already do, because it is Chapters 2 through 5 aimed at a different genre of text: say the task precisely, supply the context only you have, show an example, ask for the shape you want back, steer in rounds. The material is unfamiliar to her; not one of the habits is.

The top rung is the agentic tool. Give it a task in a sentence or two and it works by itself for a while: it edits the files, runs the result, reads the error the computer hands back, changes the code, runs it again — round after round, until the task is done or it is stuck. What the human supplies is the task at the start and the verdict at the end. What that loop actually is, mechanically, is the last page of this chapter, and every part of the answer is something you have already read.

The products sitting on these three rungs are renamed, merged, relaunched and repriced faster than a book can track, which is why this one names none of them. The rungs themselves have been stable for years, and they are what you are actually choosing between.

What Milo Actually Does All Day

Tessa spends an hour watching Milo work, expecting to see typing. She sees almost none.

He starts by writing the task: what the page must do, for whom, what it must look up, what it must not touch. Read that list again — it is the four specifics from Chapter 2, worn by a developer. The assistant produces a change. Milo reads it, line by line, and rejects about a third of what comes back. He runs the automatic checks that came with the project. One of them fails; he says precisely why the change is wrong and asks for another version. Then he reads that one too.

The ratio is the surprise of the page. Most of the characters that end up in the file were typed by the assistant. Almost none of the decisions were. Milo's day did not become less demanding when the assistant arrived; it became less typing and more judgement, which is a harder day, not an easier one.

Waymark's contracts department went through this a decade earlier with translation. Machine translation arrived, the drafts got good, and the translation bureau did not close. What changed was the ratio inside it: the machine produced the first version and the translator became a reviewer who fixed, cut and signed. The bureau's name still went on the finished document, and the liability still went with the name. The machine changed who typed. It did not change who was answerable.

Why This Chapter Is in This Book

Tessa writes no code and this chapter will not change that. It is here for three things she will do anyway, all of them this year.

She will commission software — she just did, in Chapter 8, and the next request is already forming. She will sit opposite vendors who build with these tools and quote timelines that assume them. And she will hear the sentence "AI wrote it" used both as reassurance and as an accusation, sometimes in the same meeting, and will need to know which one is warranted.

The next three pages give her that. One page on what these tools make genuinely, honestly possible. One page on the bill that arrives three weeks later when nothing else is in place. One page on the handful of habits that keep the first and prevent the second — habits she has been practising on prose since Chapter 3 without knowing they had a second home.