AI 浏览器还没捂热,给 Agent 用的浏览器已经来了:Ego LiteThe browser for AI hasn't even warmed up yet, and the browser for Agents is already here: Ego Lite


AI free token/AI免费积分

刚刚还在用 Agent 处理邮件 Mail,现在我又用 Agent 驱动浏览器获取信息了。

 

1

 

最近我在使用的一款浏览器工具叫做 Ego Lite。这款浏览器,是给 Agent 准备的。比如我在 Agent 工具(Codex)里说,

 

@ego lite 看下 mowen.cn 里,首页“正在关注”里有哪些更新,选前二十条,包括作者、文章名称(附链接),和文章的摘要。

 

AI 浏览器还没捂热,给 Agent 用的浏览器已经来了:Ego Lite

 

在此期间,Ego Lite 的人机交互界面不受任何影响,Agent 也不会打开新的浏览器界面——比如 Codex 操作 Chrome 那样,在你的浏览器里点来点去翻找内容——Agent 只是默默干活,然后给你最终结果。

 

这叫 Agent Browser。

 

2

 

之前我推荐和使用过的 AI 浏览器(Dia、Tabbit 等),是在浏览器里植入大模型,然后做大量的工程工作,让用户可以在阅读、查找、翻译、重构页面、总结和写作的时候,快速使用大模型和我们一起协作。这种浏览器当然很好用,并且是给人类使用的。

 

Ego Lite 其实是给 Agent 用的,当然人类也能用,这一点和我们之前介绍过的 QQ 邮箱团队出品的 Agent Mail 如出一辙。Agent Mail 更加超前,甚至没有人们操作的入口,只能用 Agent 操作邮件。

 

让 Agent 打开浏览器,以前的方案基本是 Computer Use,操作电脑打开浏览器,打开网页、滚动、点击、复制、整理等等。Agent 要做什么呢?保持登录态,看懂页面结构,能点击正确的按钮,必要时还得从页面脚本里找到真实数据接口。

 

Ego Lite 的做法不是在浏览器里“看网页”,它更像一个程序员坐在浏览器后面干活。页面只是入口,DOM、网络请求、前端 bundle、返回的 json、并行任务……这些才是 Agent 真正要处理的内容。

 

也就是说,浏览器本身成为了 Agent 的工作台。

 

3

 

这个方向就很有意思了。和人不一样,对于浏览器,Agent 需要的是更底层的东西。它不只要读页面,还要稳定地控制页面,不打扰用户已有窗口,必要时继承登录态,又能在自己的隔离空间里完成任务。

 

Ego Lite 是一个基于 Chromium 的浏览器,怎么做到给 Agent 用呢,还是老方法:CLI 和 Skills。

 

当我们安装 Ego Lite 的时候,系统会自动为 Agent 暴露一套 CLI 可访问的 Node.js 运行时环境。Agent 通过 ego-browser nodejs执行脚本,脚本里预置了 snapshotText、click、js、cdp、browserFetch、captureScreenshot等 helper。也就是说,Agent 在用脚本驱动一个真实浏览器:打开页面、等待加载、读取页面状态、拿语义快照、点击元素,再验证结果。

 

Ego Lite 有个挺关键的概念叫做 task space,是给 Agent 使用的隔离浏览上下文,每个 task space 有自己的标签页,同时默认继承当前用户的登录态。

 

这个设计很精妙:Agent 可以进入需要登录的网站做事,但不会直接和用户正在使用的浏览器窗口抢控制权。一次任务如果要分几轮完成,也可以复用同一个 task space,避免每次都从零开始。

 

所以这个 Ego Lite 和当前所有的 AI 浏览器都不一样,它可以正常使用,就是个 Chrome:

AI 浏览器还没捂热,给 Agent 用的浏览器已经来了:Ego Lite

不过,Ego 在浏览器旁边为 Agent 做了一套可调用的操作层:创建 task space,打开或复用标签页,读取 pageInfo,获取语义快照,点击元素,执行 JavaScript,必要的时候还会截图验证,最后关闭任务空间,打完收工。

 

比如我还可以这样:看下 CatReader 里最新更新的 20 篇文章,附上链接和摘要

 

AI 浏览器还没捂热,给 Agent 用的浏览器已经来了:Ego Lite

 

墨问的 CatReader,首页并没有直接列出最新 20 篇文章,只展示分类、文章总数、AskCat 入口和阅读面板等。传统做法可能是靠肉眼找列表,或者在页面里点来点去。Agent 的路径更像调试一个真实 Web 应用:先观察页面,再探测接口,发现同域 /api/articles返回的是前端 HTML,于是继续读前端 bundle,找到实际 API base 和文章接口形态。

 

很多网页对人是“可见” 的,对 Agent 却不一定“ 可用”。人可以看见按钮在右上角,Agent 需要知道这个按钮有没有可访问名称、是不是自定义 div、点击后状态是否改变。人可以靠经验判断一个列表按时间排序,Agent 需要拿到返回字段,确认 publishedAt、fetchedAt、feedTitle和 link。

 

Ego Lite 提供了三条路径。普通网页优先走语义工作流,用 snapshotText()拿到页面文本、按钮、链接和可点击引用,再用 click或 fillInput操作。遇到富文本、表格、地图、白板这类 DOM 和真实界面不完全一致的页面,就退到视觉工作流,用截图、坐标和键盘动作验证。需要读页面内部状态、查 DOM、调用浏览器协议时,再用 js()和 cdp()。

 

js()这点尤其有意思。它本质上接近 Chrome DevTools Protocol 里的 Runtime.evaluate,是在浏览器页面上下文里执行代码。所以刚才查 CatReader 时,它会先读页面脚本,再从前端 bundle 里找到真实 API base,最后用浏览器上下文请求文章接口,它能直接理解页面结构和前端状态。

 

4

 

Agent 浏览器最大的价值是让 Agent 拥有了一个可持续工作的前台环境。很多任务都卡在:邮箱需要登录,内部系统有权限,网页有动态加载,数据藏在前端状态里,普通 HTTP 请求拿不到东西。过去 Agent 要么让人复制粘贴,要么写脚本绕路,要么就是直接操作浏览器。

 

Agent 浏览器呢,既可以像人类一样打开页面,也可以像工程师一样检查页面。它有自己的隔离任务空间,不必抢用户正在使用的窗口;它可以复用用户登录态,又不把控制权和日常浏览混在一起;它还能在需要人工介入时交还控制权,等用户确认后继续工作。

 

这类浏览器也改变了“验证” 的含义。以前让 AI 写一段结论,验证大多靠模型自我检查,或者靠人再看一遍。现在 Agent 可以打开真实网页,执行真实操作,读取真实返回,再把结果整理出来。它可以发现页面上“广场” 按钮不是普通文本节点,也可以发现某个 API 路径实际返回 HTML。模型的判断开始被浏览器现场校准,这比空想可靠得多。

 

对于 Vibe Coding 也有很大的帮助,如果是 Web 应用,直接让 Ego Lite 理解功能验收成果,更加快捷、准确和方便。

 

5

 

普通用户对 Agent 浏览器的感知不会很强,对大众来说,AI 浏览器更有吸引力,AI 获取浏览器上下文、直接给方案、还能使用大模型重塑界面,辅助写作,等等,完全是个全新的功能。而给 Agent 用的浏览器,就像基础设施,它的价值体现在长任务里:查资料、做任务、验证页面、整理数据、测试产品、排查问题等等。

 

它为浏览器提出了另一组默认值:通过 Agent 和 Skills 理解页面、读取状态和上下文、保证安全,不打扰用户。

 

AI 浏览器解决的是人和信息之间的距离。Agent 浏览器解决的是任务和执行环境之间的距离。前者让网页更容易被理解,后者让网页更容易被操作,这么一看,两者相辅相成,颇为奇妙。

 

看起来,所有的工具,都需要为 Agent 再做一遍。

 

 

I've been using Agent to handle email in Mail, and now I'm using it to drive a browser to retrieve information.

1

Recently, I've been using a browser tool called Ego Lite. This browser is built for Agents. For example, in the Agent tool (Codex), I can say,

"@ego lite, check the 'Following' section on mowen.cn's homepage, select the top 20 updates, including authors, article names (with links), and abstracts."

The browser for AI hasn't even warmed up yet, and the browser for Agents is already here: Ego Lite

During this time, the Ego Lite UI remains completely unaffected, and the Agent doesn't open a new browser window — unlike when Codex opens Chrome, which takes over your browser and clicks around. The Agent works quietly in the background and then reports the results.

This is called an Agent Browser.

2

Previously, I recommended and used AI browsers (Dia, Tabbit, etc.), which integrate large models into the browser. They perform a lot of engineering work so that users can quickly use AI for reading, searching, translating, restructuring pages, summarizing, and writing. These browsers are very useful for humans.

Ego Lite, however, is primarily for Agents, though humans can use it too. This is similar to the Agent Mail developed by the QQ Mail team, which is even more advanced because it has no user interface; it can only be operated by an Agent.

In the past, when an Agent needed to open a browser, the typical solution was "Computer Use" — controlling the computer to open the browser, navigate, scroll, click, copy, and organize. The Agent needed to maintain the login state, understand the page structure, click the right buttons, and sometimes read the data interfaces from the page scripts.

Ego Lite's approach is different. It doesn't "read the webpage" like a human. It behaves more like a programmer working behind the browser. The webpage is just the entry point; the Agent actually deals with DOM, network requests, front-end bundles, returned JSON, and parallel tasks.

In other words, the browser itself becomes the Agent's workbench.

3

This direction is very interesting. Unlike humans, Agents need lower-level access to the browser. They need to read the page, control the page reliably, not interfere with the user's existing windows, inherit the login state when necessary, and complete tasks in their own isolated space.

Ego Lite is a browser based on Chromium. How does it make itself usable for Agents? The same old way: CLI and Skills.

When we install Ego Lite, the system automatically exposes a Node.js runtime environment that the Agent can access via CLI. The Agent executes scripts using "ego-browser nodejs", which includes pre-defined helpers like snapshotText, click, js, cdp, browserFetch, and captureScreenshot. So, the Agent is essentially scripting a real browser: opening pages, waiting for loading, reading page states, taking semantic snapshots, clicking elements, and verifying results.

A key concept in Ego Lite is the "task space", which is an isolated browsing context for the Agent. Each task space has its own tabs and inherits the current user's login state by default.

This design is brilliant. The Agent can enter a site that requires login and perform tasks without fighting the user for control over the main browser window. If a task requires multiple steps, the Agent can reuse the same task space rather than starting from scratch each time.

Because of this, Ego Lite looks and works just like a normal Chrome browser to the user:

The browser for AI hasn't even warmed up yet, and the browser for Agents is already here: Ego Lite

However, next to the browser, Ego creates a layer of operations that the Agent can call: creating task spaces, opening or reusing tabs, reading pageInfo, getting semantic snapshots, clicking elements, executing JavaScript, and capturing screenshots for verification, before finally closing the task space.

For example, I can ask: "Find the latest 20 articles in CatReader, with links and abstracts."

The browser for AI hasn't even warmed up yet, and the browser for Agents is already here: Ego Lite

The CatReader homepage doesn't list the latest 20 articles directly; it only shows categories, total articles, the AskCat entrance, and the reading panel. Traditionally, you'd have to look for the list visually or click through the page. The Agent's path is more like debugging a web app. It observes the page, probes the interfaces, discovers that the /api/articles endpoint returns HTML, digs into the front-end bundle, and finds the actual API base and the article interface structure.

Many websites are "visible" to humans but not "usable" to Agents. A human can see a button in the top right; an Agent needs to know if that button has an accessible name, if it's a custom div, and what happens when it's clicked. A human knows a list is sorted by time; an Agent needs to read the return fields (publishedAt, fetchedAt, feedTitle, link) to confirm this.

Ego Lite provides three paths. For normal web pages, it uses the semantic workflow — getting text, buttons, and links via snapshotText() and manipulating them with click or fillInput. For rich media pages (text, tables, maps, whiteboards) where the DOM doesn't match the visual interface, it switches to the visual workflow, using screenshots, coordinates, and keyboard actions. When it needs to read internal page states, check the DOM, or use the browser protocol, it uses js() and cdp().

The js() function is particularly interesting. It's essentially the Chrome DevTools Protocol's Runtime.evaluate. It runs code in the context of the browser page. So, when checking CatReader, it can read the page scripts, find the API base in the bundle, and then request the article list using the browser's own context, allowing it to directly understand the page structure and front-end state.

4

The biggest value of an Agent browser is that it gives the Agent a sustainable front-end environment. Many tasks fail because the mail requires login, the internal system has permissions, the webpage loads data dynamically, or the data is hidden in the front-end state, making it impossible to retrieve with a simple HTTP request. In the past, Agents either had to ask the user to copy-paste, write bypass scripts, or directly control the browser.

With an Agent browser, the Agent can act like a human by opening the page or like an engineer by inspecting the page. It has its own isolated task space, so it doesn't interfere with the user's work; it can use the user's login credentials without taking over the browser; and it can hand control back to the user if manual intervention is required.

This type of browser also changes the meaning of "verification". Before, verifying a model's conclusion often relied on the model's own checks or the user's review. Now, the Agent can open the real webpage, execute real actions, read the real return values, and organize the results. It can detect if a "plaza" button is just plain text or if an API endpoint actually returns HTML. The model's judgment is calibrated by the real-world results of the browser, which is much more reliable than guessing.

This is also very helpful for Vibe Coding. For web applications, Ego Lite can directly understand the acceptance criteria and verify the functionality.

5

For regular users, the impact of an Agent browser will likely go unnoticed. For the general public, AI browsers are more attractive — they offer features like direct context understanding, scheme generation, and interface reconstruction. An Agent browser, on the other hand, is infrastructure. Its value is in long-term, multi-step tasks: researching, completing workflows, verifying pages, organizing data, testing products, and troubleshooting.

It introduces a new set of defaults for the browser: understanding the page and reading states through Agents and Skills, ensuring security, and avoiding disturbances to the user.

AI browsers reduce the distance between people and information. Agent browsers reduce the distance between tasks and the execution environment. Together, they complement each other and create a powerful ecosystem.

It looks like every tool will need to be remade for Agents.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry