People are once again discovering the gap between the benchmark performance and real-world usefulness of web use / computer use agents. But there's an even bigger problem with the idea of automating shopping & travel booking.

The current experience sucks because (1) it's hard to get the user interface right and (2) deliberate enshittification. Moving from a GUI to a natural language interface makes the user experience worse, not better. It will inevitably require many rounds of back and forth to figure out what the user wants, or else risk getting things wrong. Our baseline expectation should be that the user experience will be similar to ordering over the phone, which I think most people like less than using an app with a GUI.

As for enshittification, introducing an additional intermediary (the agent developer) who will want a cut of the transaction will also only make things worse. The relationship between Doordash and restaurants is already hostile and exploitative. It's possible that the plan is to eventually cut out the intermediary and order directly from merchants. But that introduces a new problem that having to interact with a thousand different shitty merchant websites will introduce countless edge cases and new ways to fail.

There are so many potential applications of computer use / web use agents. I'm mystified by why all these developers are focusing on these doomed use cases. Let's see what happens.