Browser Use 用 Jev 做浏览器 agent 决策层
页面压成编号元素列表，Jev 一次只选一个操作，动作空间限定八个动词，订票 demo 7.1 秒出结果。
BROWSER AGENT THAT NEVER WRITES A SINGLE WORD UNLESS THE BUTTON SAYS "TYPE HERE" Every browser agent so far works the same way: screenshot in, a big model thinks out loud, coordinates or code come back This one deleted the thinking part Browser Use built it on Jev. The page gets flattened into a numbered list of elements, like [1] button Change ticket type, [2] combobox Where from, [3] combobox Where to Jev picks one operation and one element in a single request > That's the whole loop The action space is eight verbs. CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, BLOCKED Nothing outside that list is even offered to the model, so it can't invent an action that doesn't exist A small LLM only wakes up when the chosen operation is TYPE_TEXT. Text generation is the one job that actually needs generation, so that's the only place it gets paid for Result: one goal in plain English, Zurich to London one way, September 20, and the flight options on screen in 7.1 seconds with loading waits included > No site specific scripts, no prepared field strings, and the demo checks the real route and date afterward instead of trusting the agent's word The Android version follows the same rule. Jev gets a compact semantic UI state plus a short list of already valid actions, never coordinates Its Uber run: 9 actions in about 21 seconds, stopping at payment selection > The repo says outright that a completed booking isn't demonstrated The design lesson sits in that constraint: an interface an agent can operate reliably is one that offers a short list of valid next moves, not a blank canvas full below
这条是引用着另一条写的：x.com/notegoyard/status/2103135241342693883