Process

Uncle Bob — author of Clean Code, author of Clean Architecture, the man behind the SOLID principles — doesn’t read the code his agents write. Bob writes: “I’m considerably older than you. I started programming in the late ’60s. My current strategy is to not read the code written by my agents. It’s the only way to actually benefit from their throughput. Instead, I surround the agents with extreme constraints: unit tests, Gherkin tests, QA procedures, quality metrics, mutation testing, test coverage, and a bunch of others. In the end, I’m very confident in the code they produce, because it had to pass through all of my constraints and tests.” https://x.com/unclebobmartin/status/2080257779395154409 ...

Mitchell tweeted something simple: “I READ CODE…” And that was it. The internet, Twitter especially, was not ready for such a radical statement. Every time I hear something like “Claude wrote all of this, I don’t understand how it works,” I draw one conclusion: this person is incapable of debugging, full stop. If you can’t debug code, you don’t really own it. And if you don’t own it, nobody who cares about reliability or stability can trust you as a vendor. ...

Recently I pushed code to production without reading it. Looked at the title. Looked at the description. Asked an AI agent to review it. The agent wrote something or other. Out of the corner of my eye I caught a green emoji — click, click — and the request flew off to prod. It was a planned trick, to be fair. I know the codebase by heart, there are plenty of tests. A kind of Russian roulette: where you used to switch off a data center and watch what fell over, now you commit with your eyes closed. And if it falls over, you say it was a planned stress test of the AI pipeline. ...

The morning starts with going through the logs — the errors that piled up overnight. Nothing complicated. You read the errors, you go into the code, you see why a wrong parameter came in here, and over there why a parameter is missing instead. You commit. Note to self: check again in a couple of hours. At the meeting you describe what you spent your time on — and the response is an idea. Maybe you need a skill? Look: Claude Code will do all of this for you, and you just review the merge requests. We schedule the skill with cron and that’s it. ...

A task’s status updates itself, automatically, from the pull request. Sprints have been replaced by an optional weekly call. There’s no sprint planning either — you have an LLM generate a one-page design doc for the feature, and that’s the whole ceremony. Agents work the codebase autonomously. They write the tests. They review the code. They open small pull requests and merge them. Programmers are involved only in the big architectural decisions, the designs, and in giving feedback to the agents. ...