Built an agent that hill climbs YouTube thumbnails. You can't publish slop and wait for analytics to tell you it was slop. Review has to happen before you publish. But a human can't review every iteration either. That wouldn't be very agentic. So spend the human effort on teaching a judge agent up front. Calibrate the evals once, and the agent can hill-climb without you watching every step. Here's how it works: > nano-banana-pro on fal generates a thumbnail for a HubSpot-style video > A Gemini 2.5 Flash judge scores it on a 5-criterion rubric > The agent reads the judge's notes, writes the next edit, and creates another > Over 4 iterations the score climbs from 4.8 to 8.8 Then a human reviews it. Their feedback becomes 5 rule cards in the judge: > The human-encoded judge re-scores the 8.8 thumbnail to 1.2 > The agent climbs again on its own to 8.4, then 10 > The human approves An LLM judge is only as good as the expertise you give it. The judge didn't need more compute. It needed taste. Calibrate your judge up front, then let your agent hill climb. Self-verifiability is what makes coding agents so good. Marketing is more subjective though, so you need different verification methods.