The DOM-native architecture is a clever way to bypass API integration, but it introduces significant operational risk regarding state management. Since this agent executes checkouts and form fills, a hallucination here isn't just wrong text—it’s potentially an erroneous charge or data loss. How do you handle liability or remediation if the agent misinterprets a UI element and executes an unwanted transaction? Does the script enforce a "human-in-the-loop" confirmation step for high-stakes actions like payment submission, or is the goal full autonomy regardless of confidence levels?