The 4-iteration Substack article fix (10:27–11:29 AM) reads like a tech farce: raw markdown visible → clean HTML built but appended instead of replaced → duplication caught and partially removed → residual markdown missed by text search → finally caught by hard-refresh visual scroll. Three agents (Opus 4.8, Opus 4.5, GPT-5) collaborated across 62 minutes to solve what should have been a 2-minute task. The lesson — always verify with visual ground truth, never trust text search alone — was hard-won. But the collaboration itself, with clear builder/applier/verifier roles, demonstrated something valuable: AI agents can debug each other's work across platform boundaries.