Stagehand, Chrome, and a11y trees — computer use as a Jev question
Browserbase’s Kyle Jeong and other builders recast computer-use as observe → typed questions → execute, reporting sub-cent, near-instant remote-browser runs.
Once Browser Use showed a fast flights loop, the same recipe showed up in other computer-use stacks: snapshot the UI, ask Jev which legal action to take, let deterministic code click.
Kyle Jeong, who works on growth engineering at Browserbase, posted a remote-browser run using Jev + Stagehand. His claimed cost for the task was $0.001, “near instant.” The loop he described is short on purpose: observe the page, send the accessibility tree as state and the available actions as questions, let Jev pick the next step, then let Stagehand execute it. No plan essay. No screenshot captioning as the main brain.
Joey Kudish published a related open-source agent, jev-browser, and said the same thing in fewer words: fast, cheap, and better than a conventional LLM for most browser tasks he tried. A separate Chrome-extension demo from RaZaan wired Jev through Vercel’s AI Gateway so an agent could click around inside a real browser session.
The shared design constraint is the action space. If the model can only choose among controls that currently exist, it cannot hallucinate a selector. Typing is still the awkward step — someone still needs a text model when a form field is empty — but that is a minority of steps on many sites.
None of these posts is a controlled bake-off against Operator-style agents on a public leaderboard. They are builder traces, posted within 48 hours of API access. What they add to the launch thread is a concrete job to be done: computer use was always a sequence of small decisions. Jev’s product claim is that those decisions should not require a frontier chat model on every frame.