Jev Ultrafast：动态索引动作空间的浏览器智能体

TypeSafe 开发的浏览器 AI Agent 工具 Jev Ultrafast 采用动态索引动作空间架构，能够根据自然语言指令独立完成网页交互任务1。该系统支持 CLICK、TYPE_TEXT、SELECT、SCROLL_UP、SCROLL_DOWN、WAIT、DONE、BLOCKED 等操作1，并兼容 inception/mercury-2.5、Gemini、GLM、DeepSeek 等多种文本模型1。

在性能方面，Jev Ultrafast 展现出显著提升1。系统在 Google Flights 完成从 Zürich 到 London 的单程机票搜索仅需 7.1 秒1，相比之前版本中位任务时间从 9.450 秒降至 7.092 秒，性能提升 25%1。浏览器协议调用次数也大幅减少，从 1,092 次降至 101 次1。此外，Wikipedia 文章搜索任务完成时间为 2.798 秒，本地酒店搜索过滤任务为 1.896 秒1。该工具已开源，代码地址为 https://github.com/browser-use/jev-ultrafast.git[1](#source-1)。

TypeSafe has released Jev Ultrafast, a browser AI agent designed to execute web interaction tasks through natural language instructions using a dynamic, indexed action space architecture 1. The system demonstrates significant performance improvements, completing a one-way flight search from Zürich to London on Google Flights in 7.1 seconds, which represents a 25% performance gain compared to its predecessor 1. The median task completion time improved from 9.450 seconds to 7.092 seconds, while browser protocol calls were reduced from 1,092 to 101 1.

The tool supports a range of actions including CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, and BLOCKED 1. Additional benchmark tasks showcase its efficiency: Wikipedia article searches complete in 2.798 seconds, and local hotel search filtering tasks finish in 1.896 seconds 1. Jev Ultrafast is compatible with multiple text models, including Inception/Mercury-2.5, Gemini, GLM, and DeepSeek 1. The source code is available as open source at https://github.com/browser-use/jev-ultrafast.git 1.