In the real app, your agent reads the feedback JSON and revises every flagged paragraph. You review again until everything is approved.