The Prompt Isn't the Product.

Weekly requests to Uber's internal coding agents grew 9.4 times over between February and August this year, while weekly active users across the company — engineers and non-engineers both — grew seven times over. Cost per session is down 52% from its June peak. More than 70% of pull requests are now agent-authored. Those are Uber's own numbers, from its own engineering blog, and they are exactly the kind of numbers that get forwarded around a leadership team with “we should be doing this” attached.

The number is real. What it measures is not “AI replaced engineering.” It's request volume flowing through a system that engineers built, evaluate continuously, and pay for by the request — 3,600 internal agent skills, run 30,000 times a day, against routing and cost accounting someone had to design first. I watched the same gap open up at much smaller scale, on a permission check, this year.

A prompt is not a delivery process

We had two hand-rolled types answering the same question two different ways: had a permission dialog already fired. Collapsing them into one type was the kind of change a prompt handles well — mechanical, well-specified, easy to verify by reading. It compiled on the first try. Every call site type-checked.

It was also backwards. The old types answered “did it already fire?” — so call sites read guard !oldLatch() else { return }. The replacement, a type called Once, answers the opposite question — “did I get it?” — so the same idiom silently inverts: the first caller returns early, the permission dialog never appears, and the tool fails with a message about System Settings that nobody can act on. The compiler cannot catch this. Both spellings are valid Bool. Nothing about the type system, the tests we already had, or the fact that it built cleanly said anything was wrong.

What the prompt did not do

The prompt did not know that the two functions it was replacing meant opposite things by the word “yes.” That's not a knowledge gap a bigger model closes — it's a fact about this codebase, invisible from the diff, that only someone who remembered writing the original two latches would think to check. Finding it meant reading both call sites by hand, asking what “true” was supposed to mean at each one, and writing a test that claims the permission 500 times concurrently and checks that exactly one caller wins. The prompt produced correct-looking code. The engineering was noticing that “correct-looking” and “correct” were not the same claim, and building the one test that could tell them apart.

Amplifies is not the same as replaces

None of this makes the tool less useful — the collapse from two latches to one was the right call, and a prompt got me there faster than I'd have got there typing it by hand. The claim I'd push back on is the one that treats speed as the whole story. AI amplified how fast I could produce a plausible answer. It did nothing to amplify how fast I could tell a plausible answer from a correct one — that part was still mine, and it was still the part that mattered, because the failure mode here isn't a crash. It's a feature that silently stops working and reports a message that sends the next person looking in System Settings instead of at the code.

I've made a version of this argument twice already this year — first in Compiled Is Not Done, then a month later in Compiled Is Not Done: Why “It Ran” Is Not “It Works” for AI, because a compiler and a test suite both agreeing with a change is not the same claim as the change being right. The Once bug is the smallest, most literal version of that I've hit: a type that compiles, passes review, and is backwards.

The one question

Before you count a piece of AI-assisted work as done, ask who checked it and how — not whether it compiled, ran once, or looked right on the first read. If the honest answer is “it looked right,” you have a prompt, not a product yet.

Speed is the easy half. If your team is shipping AI-assisted work faster than it can verify it, the gap is a delivery process, not a model. Let’s look at where yours actually closes. Book a walkthrough.