Bob Martin, who’s been coding since the late ’60s, says he doesn’t read any code written by his AI agents. He leans into their speed by treating their output like a black box. Instead of manual review, he builds a tight testing framework around it. He layers on unit tests, Gherkin-style behavior tests, QA procedures, mutation testing and code-coverage metrics. Every piece of code has to survive that gauntlet. If it fails at any step—say, coverage dips below a set percentage or a mutation test breaks—it doesn’t get merged. That testing ensemble gives him confidence in the final product. He trusts the agents to crank out code quickly, then lets his test suite act as gatekeeper. No reading, just constraints.