BROWSER AGENT / COMMUNITY CASE
jev-ultrafast
A browser-agent experiment where a small text model handles typing while Jev picks the operation and the observed element.
A browser agent still has to produce text: search queries, form values, addresses. That typing is slow when a large model writes it token by token.
This project separates the two jobs. Jev supplies the judgment — which operation to perform and which observed element to act on — and a small, dedicated text model produces the string that gets typed.
Jev makes the per-step selection: the operation and the target element, chosen from the structured page observation. Text generation is delegated, so the judgment stays small and the typing path stays fast.
Case notes by JevCases. Repository snapshot dated ; claims not independently reproduced.
The README's approximate seven-second Google Flights example is reported by the project author.
← RETURN TO THE CASE INDEX