文档Docs
我为什么做 Agent J —— 给所有用户的一封公开信Why I built Agent J — an open letter to all our users
给所有用户的一封公开信：Agent J 从哪里来、眼下的难点，以及接下来要做的事。An open letter to everyone using Agent J: how it began, the challenges today, and what comes next.
Agent J 不是一场头脑风暴想出来的产品。它先是我自己每天在用的超级工具，后来身边的朋友想用，我才把它抽象成产品。想把来龙去脉讲清楚，就从我自己的工作方式说起。
从一个人的频道开始
宏观阿尔法频道是我用 Agent J 做出来的。所以我希望把这套使用 AI 的认知沉淀成一个产品，让更多人也能用上。
用浏览器做遥控器，只是操作层面的「术」。它最早解决的是我出差时的问题：到了香港、新加坡这些地方，没法直接用 Claude Code 或 Codex，只能翻墙，而翻墙会让账号暴露在风险里，出差越久风险越大。再加上我的工作主机后台常年跑着大量自动化，我不可能抱着它到处开会，更不想中断这些进程。于是有了一个想法：主机留在家里，浏览器当遥控器。
从二十多个工作流到一个董事长助理
另一个出发点在管理上。我习惯把人类当董事长、把 AI 当虚拟公司的 CEO，一条工作流就是一家虚拟公司。董事长只管好 CEO，具体流程让 CEO 自己去编排。
管一两条工作流时，这样完全没问题。可当工作流增加到二十多条，我发现自己的带宽不够了。于是我加了一位董事长助理，也就是 Agent J，让它去管理这二十多位虚拟 CEO。这样我就从同时盯着十几个会话，回到了只跟一个人说话。
用了一段时间，我发现这种方式特别好。我随口说的话往往零碎、不完整，但 Agent J 会把它整理成专业、完整的指令交给对应的 CEO，并且盯着他们把事做完。所以 Agent J 这个产品，就是我过去操作 AI 的经验和认知沉淀下来的结果：以前是我内部用的工具，现在成了对外的产品。
真正的难点：安装和初始化
自己一个人用的时候，环境和安装都不是问题。我是工程师出身，疫情期间又自学了编程。但当身边朋友也想用时，我把它提炼成产品、帮他们一台台装，边装边改。产品上线大约十五天，已经迭代了三十多次，其中一半都花在安装、初始化和断线修复上。对工程师来说，敲几行终端命令不算什么；可对很多人来说，连终端在哪里都找不到。
所以 Agent J 现在要做到的是：装好一次，主机就可以永远放在角落。如果是笔记本，合上盖子也不用再碰它。所有事情都在遥控器上完成，包括终端操作。最新版本已经能在遥控器的设置里点一下，就在主机的终端里执行常用指令，并把结果送回遥控器；下一个版本会直接在遥控器里内置网页终端。
安装只是第一步。真正的初始化，是在 AI 能控制的浏览器里，把自己常用的服务都登录好。很多国内的第三方服务没有 API，想把工作流程自动化，大部分环节只能在浏览器里完成。所以浏览器的配置、各种第三方 API 的申请，以及把这些 key 填进 Agent J 的密钥卡，依然会卡住很多人。这是我接下来最想解决的问题。
剑谱，而不是宝剑
Agent J 不像 OpenCode 那样自带一整套 Harness、能独立运行。我们依赖模型厂商自己的 Harness，或者 OpenCode 这类开放的 Harness，在它们上面加一层隧道，让你任何设备上的浏览器都变成遥控器，安心地遥控主机干活。
如果说模型厂商卖的是宝剑，我们卖的就是剑谱。卖宝剑的也会附一本说明书，教你怎么挥，但每家都有自己的套路，也都绑定着自家的宝剑。而用户确实需要学到更多的剑法。
关于国产大模型
对大陆用户来说还有一道门槛：很难拿到正规的 ChatGPT 和 Claude 账号，而中转账号都不稳定。上游账号是共享的，五小时用量限流分分钟就会触发，体验极差。
所以我把国内三家主力模型都接通了：Kimi、火山方舟，还有阿里百炼的千问。底层用 OpenCode 接国产模型，上层用 Agent J 遥控，就能在国内直接跑起来。说实话，我过去很少用国产模型，因为在海外用美国最好的两家模型很方便；但这段时间为了适配更多大陆用户，我测了很多国产模型，它们进步很大，比我想象的好得多。
我觉得这只是时间问题。我现在住在日本。当年日本是全世界电子产品的中心，好东西都要来日本找；可今天想在这里找点新奇好玩的电子产品，尤其是 Vibe Coding 相关的，几乎找不到。再看电动车，曾几何时听到比亚迪大家都觉得品牌不行，现在买国产电动车已经不会觉得低人一头。中国制造在一步步品牌化，大模型也会一样。
等国产模型越来越好、越来越多人愿意用的时候，翻墙和中转账号都不再是问题。我们现在布局，是希望到那一天，这套编排技术、浏览器遥控器的概念，以及「AI 董事长助理替你管几十上百条自动化和工作流」的理念，已经有了自己的位置。
接下来
我们自己的底层 Harness 已经开发好了。有了它，Agent J 就不必再绑定任何第三方应用，可以做成一个一键安装的独立 App，跟你手机电脑上所有的 App 一样。
我自己每天就是这样工作的：躺在沙发上，一边看电视，一边用手机浏览器给董事长助理安排任务、看它监督后台的工作。Agent J 本身，也是 Agent J 开发出来的。
希望它也能成为你的董事长助理。
—— Leo（王利杰）
2026-10-18
Agent J didn't begin in a brainstorming session. It was a powerful tool I used every day. Only when friends wanted to use it did I turn it into a product. To explain how that happened, I need to start with the way I work.
It started with a one-person channel
I built the MacroAlpha channel with Agent J. I wanted to turn what I had learned about working with AI into a product that more people could use.
Using a browser as a remote control is the practical part of that idea. It first solved a problem I had while travelling: in places such as Hong Kong and Singapore, I couldn't use Claude Code or Codex directly and had to use a VPN. That put my accounts at risk, and the longer the trip, the greater the risk. Meanwhile, my main work computer was always running a lot of automations in the background. I couldn't carry it from meeting to meeting, and I didn't want to interrupt those processes. So I had an idea: leave the computer at home and use a browser as the remote.
From twenty-plus workflows to a chairman's assistant
The other starting point was management. I think of the human as the chairman and the AI as the CEO of a virtual company. Each workflow is a company. The chairman manages the CEO, and the CEO figures out how to run the work.
That worked well with one or two workflows. But once I had more than twenty, I ran out of bandwidth. So I added a chairman's assistant—Agent J—to manage those twenty-plus virtual CEOs. Instead of watching a dozen conversations at once, I could go back to talking to just one assistant.
After using it for a while, I found this approach worked extremely well. My spoken instructions are often fragmented and incomplete. Agent J turns them into clear, professional briefs for the right CEOs and follows through until the work is done. That's where the product came from: years of experience and thinking about how to work with AI, distilled into something others can use. It used to be my internal tool; now it's a product.
The real challenge: installation and initial setup
When I was the only user, installation and the environment weren't a problem. I have an engineering background, and I taught myself programming during the pandemic. But when friends wanted to use it, I began turning it into a product, installing it for them one computer at a time and improving it along the way. About fifteen days after launch, we had already shipped more than thirty iterations. Half of that work went into installation, initial setup, and fixing disconnections. An engineer thinks little of typing a few terminal commands. Many people don't even know where to find the terminal.
So this is what I want Agent J to make possible: install it once, then leave the host computer in a corner. If it's a laptop, close the lid and leave it alone. Everything should happen through the remote, including terminal operations. In the latest version, you can tap a common command in the remote's settings, run it in the host computer's terminal, and get the result back on the remote. The next version will put a web terminal directly inside the remote.
Installation is only the first step. The real setup means signing into all the services you use in a browser that the AI can control. Many third-party services in mainland China have no API, so automating a workflow means doing most of it through the browser. Configuring that browser, applying for third-party API access, and adding those keys to Agent J's key card still leave many people stuck. That's the problem I most want to solve next.
The swordsmanship manual, rather than the sword
Agent J doesn't come with a complete Harness that runs independently, as OpenCode does. We rely on the model providers' own Harnesses, or an open Harness such as OpenCode. We add a tunnel on top, turning the browser on any device into a remote control so you can comfortably direct work on your host computer.
If model providers sell the swords, we sell the swordsmanship manual. A sword comes with instructions, but each maker teaches its own techniques, tied to its own blade. People want to learn more ways to use one.
Chinese models
Mainland Chinese users face another barrier: official ChatGPT and Claude accounts are hard to obtain, and intermediary accounts are unreliable. Those upstream accounts are shared, so the five-hour usage limit can kick in almost immediately. The experience is terrible.
That's why I connected three major Chinese model platforms: Kimi, Volcano Ark, and Qwen through Alibaba Bailian. With OpenCode connecting to Chinese models underneath and Agent J providing the remote on top, the system can run directly in mainland China. Honestly, I rarely used Chinese models before, because living overseas made it easy to use the two leading American providers. But while adapting the product for more mainland users, I've tested many Chinese models. They've improved enormously—much more than I expected.
I think it's a matter of time. I live in Japan now. Japan was once the world's electronics centre, the place you came to find the best products. Today, I can barely find anything new and exciting here, especially for Vibe Coding. Electric cars tell a similar story. People once thought poorly of BYD as a brand. Now, buying a Chinese electric car no longer makes anyone feel they're settling for less. Chinese manufacturing is building its own brands, step by step. I think large language models will follow the same path.
As Chinese models improve and more people choose them, VPNs and intermediary accounts will stop being a problem. We're laying the groundwork now so that, when that day comes, our orchestration technology, the browser remote, and the idea of an AI chairman's assistant managing dozens or hundreds of automations and workflows will already have a place.
What comes next
We've already built our own underlying Harness. With it, Agent J won't need to depend on a third-party application. It can become a standalone app with a one-click install, like the other apps on your phone and computer.
This is how I work every day: lying on the sofa, watching television, and using my phone browser to assign tasks to my chairman's assistant and watch it supervise the work in the background. Agent J itself was built by Agent J.
I hope it can become your chairman's assistant too.
— Leo (Wang Lijie)
2026-10-18