Why trust an AI reviewer if you don't trust the agent it checks?

A user questions whether Codex and Claude Code's auto-approval offers reassurance if permission comes from another agent whose judgment you may not trust. In a follow-up, they describe studying this through a model of the user and reviewer agents' utility functions—what each seeks to maximize. In that model, the acting agent repeatedly proposes actions, while reviewers compare them with a baseline and vote to approve or deny based on perceived utility.