Correcting AI mistakes one prompt at a time is exhausting and doesn’t scale. The better approach: let the agent fix its own mistakes and update its own configuration.