The author of Clean Code spent decades reminding us that reading code is most of the job. This week he said he stopped reading what his agents write, and that confidence comes from somewhere else: a gauntlet of constraints made of tests, metrics, mutation testing and coverage. I agree, and I explain with the caveats that matter: the gauntlet has an anatomy, and the human's role shifts from inspector to designer.

The agent writes in twenty minutes, you spend the afternoon reviewing: generating code got cheap, verifying stayed expensive, and you became the bottleneck. After more than a year refining a guideline for developing with agents, this is the philosophy and practice of what I call Spec-Anchored Agentic Development: one spec per capability as the anchor of verification, context in three layers, the machine giving the first no before any human review, and autonomy earned per class, with evidence instead of faith.