Uncle Bob: "I don't read the code my agents write"

What's this about?

Robert C. Martin — author of *Clean Code*, coding since the late 1960s, and the profession's most famous advocate of reading code carefully — says his current strategy is to *never* read what his AI agents write. Humans read too slowly; reviewing line-by-line erases the productivity gain. Instead he surrounds agents with "extreme constraints": unit tests, Gherkin behavior tests, mutation testing, coverage thresholds, quality metrics, and a four-agent pipeline (specification → code → refactor → architecture review).

What exactly is "mutation testing"?

A way to test your tests: deliberately inject small bugs ("mutations") into the code and check the test suite catches them. If it doesn't, your tests are decorative. It's the kind of gate that lets you trust code you never look at — the tests are being audited, not the code.

What's the community saying?

1. The whiplash: days ago the debate was "how carefully should you review AI code?" — the Clean Code author just answered "not at all, verify differently."
2. Supporters call this the honest endgame: humans move from reading code to engineering the verification harness around it.
3. Critics: tests written or influenced by the same agents are a circular guarantee, and metrics like coverage are gameable.
4. Meta-observation: this is exactly the "harness engineering" philosophy that today's "Why Software Factories Fail" essay says is *not enough* — the two pieces read as a debate.