The idea
Gregor Zunic demonstrates a browser agent that turns the current DOM into a fresh set of available actions on every step. A smaller language model handles typing when a choice alone is not enough.
Inside the decision loop
Page state + available actions
Choose the next action
Browser executes the selection
What to take from it
The useful separation is between choosing an action and writing arbitrary text. Keep the action list close to the current page so a selection cannot refer to an element that has disappeared.
The author reports a short flight-search run. A single recorded task does not establish reliability across booking sites, pop-ups, or changed page layouts.
Follow the original work
This case is an editorial interpretation of a public community demonstration. It has not been reproduced or benchmarked by this publication.
Original post by @gregpr07