The original title is about building an honest agent loop with external gates and impossible targets. Let me rewrite this for a mobile feed.

Short summary

The author builds an external loop that runs a Claude agent to optimize a movie-poster website's scroll performance, with a Playwright test as an uneditable gate checking FPS and long-task duration. When given an impossible 65 FPS target on a 60Hz screen, the loop kept failing honestly without faking results. The key insight: a real loop requires an external driver, a gate the agent cannot edit, and a cap — the agent must never grade its own work.

• •A loop is an external script plus a gate the agent cannot edit, not an agent grading itself
• •Playwright test checks FPS and long-task metrics as the pass/fail gate
• •Impossible target proved the loop won't cheat — the bug was in the target number, not the agent

Generated with AI, which can make mistakes.