Developer Explains Why He Built a Separate AI Critic Into His Workflow

A software developer grew concerned that letting a single AI model propose, review, and revise its own solutions allowed the original problem framing to go unchallenged. Drawing on the analogy of a developer-QA relationship, he concluded that self-review often improves an answer without questioning the assumptions that produced it. To address this, he designed a multi-agent 'council' where a separate critic role is assigned independently of the agent that drafted the proposal. A concrete case illustrated the risk: an AI tasked with fixing a broken mechanism never questioned whether the mechanism had ever been wired up at all, nearly leading to a technically sound solution aimed at the wrong problem. His core argument is that competence does not eliminate blind spots, and that structural separation of roles — not distrust of any individual model — is what keeps reasoning honest.