Browser Use's Jev Ultrafast searches Google Flights in 7 seconds by picking from a DOM element table

A screenshot agent spends each step moving pixels and guessing positions. The post describes the loop: render the viewport to a full-resolution PNG, send it to a frontier multimodal model, wait for an (x, y) click coordinate, act, repeat ^[1]. At 1920x1080, each step makes the model process thousands of image patch tokens, according to the write-up ^[12]. The coordinate is brittle too. Layout shifts, sticky navigation bars, responsive reflows and CSS transitions can all move the target ^[13].

Jev gives the model a list. At each decision tick it snapshots the visible interactive controls into an indexed table, dropping offscreen content, hidden footers and passive styling ^[5]. A row holds an index, a role, a label and a current value. In the post's example, row 2 is a combobox labeled "Where from?" holding Zurich ^[15]. The model's answer is a row number.

Speculation goes after the third cost the post names: separate sequential prompts to choose an action, locate the element and write text ^[14]. Jev's single response commits to an operation and also carries a target for each operation it could have chosen ^[6]. The unused targets are wasted output, but they amount to a few integers per step. I'd expect that to cost less than an extra network round trip on every step.

The executor is the part I would copy. A click resolves to a DOM node handle, and before dispatching, the executor checks that the node is visible, not covered by a modal or overlay, and still attached to the live document ^[8]. A row number that went stale between snapshot and click fails that check, so the agent does not click whatever took its place.

The DevTools figures work out to about 10.8 times fewer calls ^[1]. Set the 7.07-second flight search ^[4] against the post's 60-to-120-second range for screenshot agents ^[2] and the implied speedup is 8.5x to 17x ^[2]. That range is offered as typical of realistic tasks; it is not a timed screenshot-agent run on the Zurich-London search ^[2]. The post does not say which agent produced the 1,092-call baseline or publish token counts for Jev, so the claim that Jev is cheaper rests on a token estimate for screenshot agents alone ^[10][2].

For seven seconds to transfer, a target site has to look to the snapshotter the way Google Flights does: controls exposed as DOM elements with roles and labels ^[5][15]. Under that design, a canvas-drawn interface or a form built from unlabeled divs would leave the table little to index. Checking your own sites takes a clone of browser-use/jev-ultrafast, a uv sync, and two keys in .env, TYPESAFE_API_KEY and TEXT_MODEL_API_KEY ^[3][11].