• Verification gives an agent "hands and eyes": it can run the application, interact with it as a user, debug it, and inspect traces or snapshots. Without access to the result, it cannot reliably iterate on its work. 16:03 17:27
  • In an agent loop, verification is the part that lets the agent compare its result with a target and try again. Poteto used this for performance work and describes "hill climbing" as repeatedly improving a result against a rubric or score. 17:55 18:33
  • His verification skill includes a custom CLI that wraps deterministic actions, including Playwright and Chrome DevTools Protocol calls. This prevents every agent from rebuilding its own verification scripts, which had wasted context, time, and effort. 19:05 22:01 23:02
  • Poteto separates work that requires judgment from mechanical work. Scripts and CLIs should handle repeatable transformations or measurements; agents should focus on tasks that need interpretation and decisions. He applies the same idea to migrations, using codemods and AST transformations for mechanical changes. 20:49 21:15 23:26 24:21
  • A skill can then stay small: it explains when and how to use the tools, while code provides consistent execution. 21:43 22:27