2 a.m., and nobody knows what this code does

It's 2 a.m. and checkout is down.

The on-call engineer pulls up the failing service and starts reading. The culprit is a 400-line function, merged three weeks ago with a thumbs-up emoji as the review. Nobody on the team wrote it. An agent did, in about ninety seconds, and at the time everyone thought that was pretty cool.

It's a lot less cool now.

I keep coming back to that scene because of one study. In 2025, METR ran a proper randomised trial with experienced open-source developers. With AI tools, they took 19% longer to finish real tasks. Afterwards, they guessed AI had made them about 20% faster.

(To be fair, that was early-2025 tooling, and METR has since said the number shouldn't be read as today's. The tools are better now. But that gap between how fast it feels and how fast it is? I don't think that's gone anywhere.)

For the record, I'm not anti-AI. I use it every day and I'm shipping more than I ever have.…