HOST: If I handle bookings and forms at work, why should I care about this study?
EXPERT: It could help you see which steps are hard to hand to a browser agent. The authors tested agents on real websites, but not as a workplace service.
HOST: So what makes a form task harder than finding a page?
EXPERT: It's really about carrying the right details through several steps. Think of preparing an appointment booking. Finding the form doesn't mean the requested details are in place.
HOST: Did the study let agents actually send those bookings?
EXPERT: For each task, the setup blocked the annotated final web request before it could commit the irreversible action. The authors say that safeguard is limited to those task endpoints.
HOST: Then what counted as passing?
EXPERT: So meeting the task's endpoint just means you get to whatever was defined as passing. Maybe that's reaching the block request with the right info, or just stopping at a permitted point once all the earlier steps are done.
HOST: What did they find, and how much can I read into it?
EXPERT: In the shared browser setup, Claude Sonnet 4.6 had the highest pass rate at about 33.3 percent of the 153 tasks. Sites do change, and that score comes from a model-based judge. The practical lesson is to check the recorded steps in the endpoint rather than assuming that just reaching a page means the job is done.