Browser-use ships Jev Ultrafast, a browser agent that finished a Google Flights search in 7.1 seconds

Browser-use published Jev Ultrafast on 16 September, a browser agent that runs on TypeSafe's Jev model and finished a Zurich to London flight search on Google Flights in 7.1 seconds. The repository reached 551 stars in the first fifteen hours and was posted to Hacker News overnight. The pitch is that operation and target are decided in one request, not two, and that the agent reads a structured element table rather than screenshots.

What the loop does

Every observation produces a numbered element table, one line per interactive control with role and label. The model chooses one of eight operations (CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, BLOCKED) and one target from that table, in a single request that speculatively answers all three target heads at once. Only the target that matches the chosen operation executes, and text generation runs as a second small-model call only when the operation is TYPE_TEXT. Screenshots are not in the default loop.

The measurement

Browser-use reports that in six alternating runs of the same Google Flights task with identical models and settings, median task time went from 9.45 seconds to 7.09 seconds, a 25 percent reduction, and median browser protocol calls went from 1,092 to 101. Both versions passed 3 out of 3 runs. The company states clearly that this is three repeats of one task on one browser profile, not a general reliability benchmark. The same policy opened a Wikipedia article in 2.798 seconds and passed a local hotel filter task in 1.896 seconds.

What is not covered

The DOM reader handles common HTML and ARIA controls, not the full accessible-name specification. Shadow roots, iframes, canvas, file uploads, pop-up tabs, nested scrolling and arbitrary keyboard widgets are outside this first release. A DONE choice still needs independent verification: the Flights example checks the one-way setting, both cities, the date and visible results before it accepts the run. Model output never becomes selectors, coordinates, shell commands or executable JavaScript, and text-helper output must parse as a small JSON object before it is typed.

The wall-clock number is the reason to look at this. Browser agents commonly spend seconds on each read-observe-decide-act cycle, and a five-page task adds up to a minute even when nothing goes wrong. Cutting that in half by removing a round trip changes what a reader can plan around: a flow that used to time out in a UI test now finishes inside the timeout, and an agent that runs several tasks in a row finishes work in the same session instead of asking for another.

Source

• browser-use, jev-ultrafast on GitHub, 16 September 2026.
• Hacker News, discussion.