Bo

Isoform 1.0 is out. Own your intelligence. Then rent it out. First, it learns you. Isoform watches how you work on your Mac and writes a portrait of you: how you plan, whose email gets a reply within the hour, which afternoons don't move. You can read it. It sits on your disk. Then it works ahead. The morning brief is ready before you sit down. The reply to that client is drafted the way you'd write it. You press Send. Then it works as you. That's Agent of You. Once Isoform knows how you work, other people can hire an agent that works the way you do, and you get paid by the minute. Reservations open today. Hiring comes next. Open source. Runs on the Claude Code or Codex subscription you already have, or a local model. Screenshots and audio are stored on your Mac. Your AI gets only what a task needs, and nothing goes out on your behalf without your OK. Download at isoform.ai. Reserve your agent's name at isoform.ai/agents.

Isoform 一直是产品公司。前两年我们把产品当服务卖。说说为什么。 我们给 PE 控股的中型公司做定制软件。AI agent 加人类 tech lead。12 周的集成项目压到 7 个工作日，工程人头减一半。管用，有了真实收入。我们管它叫 service as software，这是我们的 go-to-market。 每个项目教的都是同一课。代码从来不是难的部分。难的是懂人：他们怎么命名，之前试过什么，哪个决定在一个没人记录的会上已经定了。底下那个产品一直在做的，就是把这些抓下来。 所以工具越做越好。观察工程师，问他为什么这么改，记住，下个项目用。工具成了价值本身，外面包着的服务成了不能规模化的那部分。 如果瓶颈是懂人，为什么要通过我们的工程师，一家公司一家公司地解决？做给每个人，放在他自己的机器上，让他自己拥有。 这就是现在的 Isoform。服务是 go-to-market，产品才是目的。

Isoform was always a product company. For two years we sold the product as a service. Here's why. We built bespoke software for PE-backed mid-market companies. AI agents plus human tech leads. We cut 12-week integration projects to about 7 business days, at half the engineering headcount. It worked. We got to real revenue. We called it service as software, and it was our go-to-market. Every project taught the same lesson. The code was never the hard part. The hard part was understanding the people: how they named things, what they'd tried before, which decision was already made in a meeting nobody wrote down. The product underneath was always about capturing that. So the tooling kept getting better. Observe the engineer, ask why they made that edit, remember it for the next project. The tooling became the value. The service around it was the part that didn't scale. If understanding the person is the bottleneck, why solve it one company at a time through our engineers? Build it for every person, put it on their own machine, and let them own it. That's Isoform now. The service was the go-to-market. The product is the point.

Isoform 1.0 是一个人加一群 agent 做出来的。说说怎么搭的。 两个长期跑着的 session，两个活。hub 是一个 Claude Code session：用我的话维护 backlog，把活派给一群 worker（Claude Code、我 Mac 上的 Codex、Codex Cloud、一个 Kimi 队列、一个 app runner），review 并试合并 PR，然后用固定格式汇报：进行中、下一步、提议、阻塞、问题。 design session 是定产品决策的地方。它的产出是一块定好的板子。design session 从不下建造命令，hub 从不跟设计争。 这两个是故意分开的。一个 agent 既决定做什么又决定怎么做，在同一个 context 里，设计就会往这周便宜的方向弯。分开之后设计看不到成本，就不会退缩。 稀缺的不是 token，不是工程时间，是我的决定。所以 hub 的汇报是一个决策队列，推荐答案已经填好，我回一个字。它推荐 yes 的时候，空闲的 worker 在我说之前就开始建了。 session 是无状态的，仓库才是状态。杀掉一个 session、重启、换底下的模型，它会重新收敛。所以我的 Mac 睡着的时候云上的 lane 还在跑。 hub 是一个 chief of staff。它在做的 Isoform，也是一个 chief of staff。我先把这个论点在自己身上跑了一遍。 长版本：bozhao.substack.com

The iPhone Duo makes the most sense for people who sit in an office. You fold it when you travel. The rest of the time it stays unfolded on your desk, and you get the bigger screen.

Isoform 开源。说说为什么。 这个产品全部的承诺就是：它学到的关于你的东西是你的。画像、截图和音频都留在你的 Mac 上，没有你点头，什么都不会替你发出去。这种话不能靠公司保证，你得能自己看代码。所以代码是公开的。 第二个原因：我们建在 Claude Code 和 Codex 之上，不是取代它们。Isoform 是懂你的那一层，跑在你已经在付钱的 agent 上面。这样一层东西，你应该能看、能 fork、能换任何模型，包括本地模型。 第三个原因：开源是我们会的建法。Yetone 写了 alma 和 magpie，更早写了 Avante.nvim；我做过 BentoML。代码公开之后会发生什么我们见过：有人修，有人找问题，有人信。 不开源的部分：别人租你的 agent 时走的那个市场。它走我们的网关，那是生意。懂你的那部分是你的，给你付钱的那部分是我们的。 github.com/yubozhao/isoform

Isoform 1.0 was built by one person and a fleet of agents. Here's the setup. Two long-running sessions, two jobs. The hub is a Claude Code session that keeps the backlog in my words, dispatches work to a fleet (Claude Code, Codex lanes on my Mac, Codex Cloud, a Kimi queue, an app runner), reviews and trial-merges the PRs, and reports back in one fixed shape: In progress, Next, Proposed, Blocked, Questions. The design session is where product decisions get settled. Its output is a decided board. The design session never issues a build order. The hub never argues with a design. Those two are split on purpose. One agent deciding what to build and how to build it, in one context, bends the design toward what's cheap this week. Split them and the design never sees the cost, so it can't flinch. The scarce resource isn't tokens or engineering hours. It's my decisions. So the hub's report is a decision queue with the recommended answer filled in. My answer is one word. When it recommends yes, an idle worker starts building before I say it. Sessions are stateless. The repo is the state. Kill a session, restart it, swap the model under it, and it reconverges. That's why the cloud lanes keep going while my Mac sleeps. The hub is a chief of staff. Isoform, the app it's building, is a chief of staff. I ran the thesis on myself first. Longer version: bozhao.substack.com

Isoform is open source. Here's why. The product's whole promise is that what it learns about you stays yours: the portrait, your screenshots and your audio stay on your Mac, and nothing goes out on your behalf without your OK. You cannot take a company's word for that. You have to be able to read the code. So the code is public. The second reason: we build on Claude Code and Codex instead of replacing them. Isoform is the layer that knows you, running on top of the agent you already pay for. A layer like that should be something you can inspect, fork, and run with any model you want, including a local one. The third reason is that open source is how I know how to build. Yetone, our CTO, wrote alma and magpie, and Avante.nvim before that. I co-founded BentoML. We've seen what happens when people can read the code: they fix things, they find things, they trust it. What's not open: the marketplace where people rent an agent version of you. That runs through our gateway, that's the business. The thing that knows you is yours. The thing that pays you is ours. github.com/yubozhao/isoform

每个人都应该有一个自己的 agent 版本。说说这是什么意思，以及为什么必须这么做。 所有 AI agent 的失败方式都一样。不是模型不行，是它不懂你。写代码便宜了，写 prompt 也便宜了。贵的是知道你到底想要什么：你怎么命名、你信谁、对你来说"做完"是什么样。这是瓶颈，新模型解决不了。 了解一个人最好的办法是跟着他。所以 Isoform 看你工作，看不明白的时候问一个问题。几周之后，它有了一张你怎么做事的画像。 这张画像是 AI 时代你最值钱的东西，也是最敏感的东西。所以它在你的硬盘上，加密，是你的。不是我们的。 新的机会来了。因为它是你的，你可以把它租出去。Agent of You 是一个有你的脸、你的声音、按你的方式做事的 agent。有人需要按你的方式审一份合同、做一个 deck、看一个方案，他们雇你的 agent。活在你的机器上干，你按分钟收钱。想审一下再交，可以；不想审，也可以。 他们永远拿不到的是你的 know-how。租的人看到的是活，不是画像。用哪个模型你定，价格你定。 你的专业替别人干活，你去干别的。这就是我们在做的事。isoform.ai/agents

I was wrong about vibe coding. In March 2025 I wrote that AI would make software fit each person. I have polio and need custom shoes, so I pictured software the same way: someone builds it to fit you. People wanted the fit. They didn't wait for anyone to build it. They ask the model and make it themselves. People changed faster than the software did. My mistake: I judged what AI could do and held human behavior constant. Those two move together. What the model still doesn't have is the person. That's where the work is now. bozhao.substack.com/p/polio-…

Everyone deserves an agentic version of themselves. Here's what that means and why it has to be built this way. Every AI agent fails the same way. Not because the model is weak. Because it doesn't understand you. Writing code is cheap. Writing the prompt is cheap. Knowing what you meant, how you name things, who you trust, what "done" looks like for you: that's the bottleneck, and no model release fixes it. The best way to understand a person is to shadow them. So Isoform watches you work, and when watching isn't enough, it asks questions. Over a few weeks it has a portrait of how you do things. That portrait is the most valuable thing you own in the AI era. It's also the most sensitive. So it lives on your disk, encrypted, and it's yours. Not ours. Now the part that's new. Because you own it, you can rent it out. Agent of You is an agent with your face and your voice that works the way you do. Someone needs a contract reviewed the way you review contracts, a deck built the way you build decks, a plan checked by someone who thinks like you. They hire your agent. It does the job on your machine. You get paid by the minute. You review the work if you want to, or don't. What they never get is your know-how. The renter sees the work, not the portrait. You pick which model does the work. You set the rate. Your expertise, working for other people, while you do something else. That's what we're building. Reserve your agent's name now at isoform.ai/agents.

awesome！

今天先说人。 @yetone 这几年是 CTO，写了 alma 和 magpie，也写了 Isoform 的很多代码。还有 @woai3c123 、@heyshiqi ，和其他几位。 两年的客户项目，无数个晚上，无数次争论这东西到底该长什么样。今天发出去的，是我们一起找到的答案。谢谢。

Isoform 1.0 发布了。 拥有你自己的智能。然后租出去。 第一步，它学你。Isoform 在你的 Mac 上看你怎么工作，写出一份你的画像：你怎么做计划，谁的邮件你一小时内回，哪几个下午雷打不动。你能读到它，它就在你的硬盘上。 第二步，它走在你前面。你坐下之前，早上的简报已经好了。给客户的回复按你的写法起好了草稿。你只要点发送。 第三步，它替你干活。这就是 Agent of You。Isoform 学会你怎么工作之后，别人可以雇一个像你一样工作的 agent，按分钟给你付钱。今天开放占名，雇佣下一步上线。 开源。用你已经有的 Claude Code 或 Codex 订阅就能跑，也可以用本地模型。截图和音频存在你的 Mac 上，你的 AI 只拿到一件事需要的内容；没有你同意，什么都不会替你发出去。 下载：isoform.ai。先把你的 agent 名字占了：isoform.ai/agents

Lfg @RobSandIA