Video unavailable. Open the source below.
NewJev Browser by Ying-Kai Liao
A browser agent lets an LLM set the goal while Jev chooses page actions and checks progress in a Playwright loop.
Video unavailable. Open the source below.
NewA browser agent lets an LLM set the goal while Jev chooses page actions and checks progress in a Playwright loop.
For each browser step, it sends a page snapshot and task state containing the goal, supplied values, recent action history, and page changes. Jev returns Noul judgments about completion, blockers, errors, login, and irreversibility, plus Choice answers for action, target element, and sometimes a value. Code reconciles those answers, applies confidence and safety rules, and uses Playwright to act; the caller receives a status it can handle or verify.
Does the page show that the task goal has been achieved?
Current page snapshot and task goal, with visible elements, supplied values, recent action history, and changes since the last action.
What is the next action toward the task goal on this page, given the recent actions?
Current page snapshot and task goal, with visible elements, supplied values, recent action history, and changes since the last action.
Wait for network and DOM quiet, then collect the page's visible text, metrics, dialogs, and elements.
Send the snapshot and task with history and changes; ask typed questions about status, action, and target.
Choose a compatible target and value, or gather a follow-up Choice for a dropdown option, key, or drag destination.
Return statuses such as needs_login, needs_confirmation, ambiguous, or likely_done when the evidence or confidence calls for handoff.
Perform the Playwright action and loop with an updated page until done, stuck, blocked, or out of actions.
The design splits browser work into typed, bounded decisions and turns uncertainty into explicit statuses. It also narrows target selection in two stages on large pages and checks for irreversible actions before proceeding.