RPA has always had a durability problem — it operates on structure rather than meaning, so layout changes break it. A model-based browser agent that understands what a checkout button is semantically — regardless of what the HTML looks like — sidesteps that problem in a useful way. What I am more curious about is the failure handling: does the agent recognize when it is about to take an irreversible action and create a confirmation step, or does it treat all actions equivalently? That edge case design seems like where the real product differentiation will happen.