Jev 浏览器 agent 搜 Google Flights 提速到 7 秒
把页面读成编号元素表,Jev 只从里面挑一个操作,调用数从 1092 降到 101,中位耗时 9.45s→7.09s,仓库开源。
A BROWSER AGENT SEARCHED GOOGLE FLIGHTS ZURICH → LONDON IN 7.1 SECONDS It was given one goal: search Google Flights for a flight from Zurich to London, with no site-specific scripts or screenshots The page gets read as a numbered element table, JEV picks one operation from there, click, type, select, scroll, wait, done or blocked It picks the element in the same request, so there’s only one round trip A small LLM only writes text when the operation is type_text The model never writes a selector, coordinate or JS, it only picks from what’s already on the page The same policy opened a Wikipedia article in 2.8s 9.45s → 7.09s median 1,092 browser calls → 101 The main difference comes from reducing how much the model has to do Fewer calls, smaller choices Worth stealing even if you never build browser agents Give the model a list of possible actions and make it pick instead of asking it to write the action itself Repo is open source, link below
这条是引用着另一条写的:x.com/noisyb0y1/status/2102309802072272956