Different perspectives can reveal hidden failure
A reviewer agent can catch missing requirements. A debate pair can explore opposing interpretations. An adversarial reviewer can test security assumptions. These patterns help when one pass is too trusting.
But disagreement is not automatically truth. You still need a judge: criteria, scoring, tests, or human decision.
Make the critic specific
A critic that says “make it better” is noise. A critic that checks factual support, policy compliance, edge cases, and acceptance criteria is useful.
For security, ask the adversary to find exploit paths, then ask the defender to patch or reject them with evidence.