Uncle Bob stopped reading his agents' code

Robert C. Martin no longer reads the code his agents write, he says. That's a bet: automated gates catch what a reviewer's eyes used to. Every team running agents at scale is making that bet now, whether they admit it or not.

In a Hacker News post, Martin laid out the replacement: unit tests, gherkin tests, QA procedures, quality metrics, mutation testing, and coverage gates. Run agent output through that, he says, and you get "very high confidence" without ever opening the file.

Grady Booch disagrees. He argues metrics can't catch the vulnerabilities and dead code an experienced engineer spots on sight, just by reading.

The irony is hard to miss. Clean Code, the book that made Martin famous, argued code gets written once and read dozens of times. Now its author says skip the reading and trust the gates.

A green test suite proves the code does what the tests check for. It doesn't prove the code belongs where it was put, or that it isn't duplicating logic three files away.

I've hit that gap myself. Code passed every gate and was still wrong at the architecture level in a way no test encoded.

Martin's bet only works if the gauntlet is genuinely thorough, mutation testing included. I'd bet most teams skip that part and lean on green unit tests alone.