The verification principle

Out of that comes the rule that should govern every deployment, and it echoes Chapter 13: agent output requires verification proportional to its risk. A green test suite is necessary, not sufficient. A human must own review for anything touching architecture, security, or money. The guide's sharp warning: Codex shifts effort from writing to reviewing, but that only saves time if your tests and review are strong enough to catch a confident, fluent mistake. "Weak verification turns an agent into a liability multiplier." If you take one sentence from this chapter into production, take that one.