A browser Agent that uses Jev to choose actions and page controls, calling a text model only when input is needed.
Chooses an action and its matching DOM element in one request; a text model generates input text.
Separates UI choices, text generation, and browser execution so each step can be inspected.
The author reports a median of about 7.1 seconds on one Google Flights task, with three runs per setup. Initial navigation is excluded; this is not a general Agent benchmark and was not independently retested.
Separate selecting a browser control from composing text to type into it.
Material analysed：
model.py validates the chosen operation and the matching target head. Unused target heads cannot trigger actions; this is an important distinction from a model returning an arbitrary executable instruction.
Inspect the supporting source ↗The performance document excludes initial navigation and describes repeated runs of one Flights task. It does not establish a general browser-agent success rate.
Inspect the supporting source ↗Useful when the page exposes meaningful DOM controls and you want to inspect operation and target probabilities separately. Screenshot-only interfaces and free-form writing require other paths.
Read action_space, choose and validate_choice in the pinned model.py. Follow CLICK and TYPE_TEXT separately; confirm that an unknown target or an inconsistent distribution raises before execution. The README lists offline guard checks separately from paid live examples.