Kane CLI runs your flow in a real browser. When it fails, the agent reads Kane's own evidence, rewrites the objective, and tries again — the loop plays out below, not just a pass/fail badge.