Jev Ultrafast
22kA browser agent from the browser-use team where each step is decided not by an LLM but by the Jev model: which action, on which element. A Google Flights search in 7.1 seconds, MIT.
We have not run it ourselves: described from the docs and the repository.
In numbers
What it is
A typical browser agent sends a screenshot or a page tree to an LLM at every step and waits for it to write out where to click. Jev Ultrafast, a new repository from the browser-use team, works differently. The page becomes a numbered table of elements, and the TypeSafe Jev model picks what to do and on which element in a single request. A generative model only steps in when something has to be typed.
The difference from browser-use, its neighbour in this catalogue, is the approach. There the agent is general-purpose and thinks with a large model at every step; here it is a narrow, fast loop where the LLM barely takes part. It is the same move we covered in our article on Jev: binary and selection decisions are taken off the generative model, and the agent gets faster for it.
How it works
CLICK, TYPE_TEXT, SELECT, scrolling, WAIT, DONE) and the element number in a single request.inception/mercury-2.5 via OpenRouter.Agent library for your own tasks.
Comments