L8 工程师的 Agentic Engineering 系统：只把判断留给人

I just have a frameless terminal window here. Uh this is using uh western. I I like western because it's really highly customizable. I can just like change everything about it. Uh what I have here is a herder session. Herder is kind of like a modern version of T-Max. I was using T-Max for like over 10 years and only recently discovered Herder and I just really like it. It's a more modern take how you manage multiple sessions in your terminal window. 我这里是一个无边框 terminal window，用的是 WezTerm。我喜欢 WezTerm，因为它高度可定制，几乎什么都能改。这里运行的是一个 Herder session。 Herder 有点像现代版 tmux。我用了十多年 tmux，直到最近才发现 Herder，非常喜欢。它用更现代的方式管理 terminal window 中的多个 sessions。

This is running on my Mac Mini and I can connect to this same session from my phone. Um, so I I can take my phone, I do a SSH connection, I can get this exact same terminal window and that's a good thing about Herder uh which is that it understands what agents are. In order to get more work done, people have to manage multiple uh parallel agent sessions. Yeah. Right. Because one agent uh can take a while to get work done. Uh so people start to spin up more sessions 它运行在… …我的 Mac Mini 上，我也能从手机连接到同一个 session。拿出手机，通过 SSH connection，就能看到完全相同的 terminal window。 Herder 的一个优点是，它理解 agents 是什么。为了完成更多工作，人们必须管理多个 parallel agent sessions，因为一个 agent 完成任务可能需要一段时间，于是大家会启动更多 sessions。

主持人 and sometimes I see people have like 20 30 sessions but then I kind of like uh was going crazy because uh 20 sessions you have to keep that in your head right. Exactly. Yeah. So that pain caused me to develop first mate because I I I just don't think that's going to be the end game. Like I don't want to spend all day just juggling between the tabs right and remembering what was what. 我有时会看到别人同时开 20、30 个 sessions，但这快把我逼疯了，因为你必须把 20 个 session 全记在脑子里。主持人：没错。 Kun：这个痛点促使我开发 FirstMate，因为我不认为这会是最终形态。我不想整天在 tabs 之间来回切换、记住每个 tab 是什么。

I developed first mate because I think uh the agents as they become more capable should be able to juggle all those things for me and I started experimenting with this setup since a few months ago and it actually works. Uh so now this is like the only agent I talk to most of the time. Right. Kun so you were elite engineer at some elite companies like Meta, Microsoft, Atlasian. What does your current AI coding setup look like? 我开发 FirstMate，是因为随着 agents 变得更强，它们应该能替我协调这一切。我几个月前开始试验这套 setup，结果真的可行。现在多数时候，它是我唯一直接交谈的 agent。主持人：Kun，你曾在 Meta、Microsoft、Atlassian 等一流公司担任高级工程师。你现在的 AI coding setup 是什么样？

I have been using AI uh to pretty much like write everything I want to build. I very rarely would manually write any code now. And when did that change by the way? I think it was a gradual process. Kun：我已经用 AI 编写几乎所有想构建的东西。现在我极少手写 code。主持人：这个变化是什么时候发生的？ Kun：我认为是一个渐进过程。

Uh so initially I think 3 years ago we had uh GitHub copilot right code completion. uh and we started with just like taking uh the code suggestions one line after uh another and then it evolved it uh started to give us like multiple lines a whole function at a time right and then I think uh like one inflection point that changed that status quo was when sonnet 3.5 v2 was introduced that was a gamechanging moment uh so that was the first time I think an agent can take a task and start to do things for us right and come back with like a complete set of results So I started working with agents around that time. 大约三年前，我们有 GitHub Copilot，也就是 code completion。一开始只是逐行接受 code suggestion，后来它能一次给出多行，甚至整段 function。我认为其中一个 inflection point… …是 Sonnet 3.5 v2 发布。那是改变游戏规则的时刻，也是我第一次认为 agent 可以接下一个 task，替我们采取行动，最后带着完整结果回来。我从那时开始与 agents 合作。

It was also very interesting because I was working at Atlassian at the time and my job was to develop coding agents. So I played with the model a lot. I tried various kind of models uh starting from GPT 3.5 it was just not working asking it to edit a file. It's already a lot of trouble. GP4 was a little bit better but still not quite there. It was really solid 3.5 V2 that's really changed the game. So that was another I think inflection point. 有趣的是，我当时正好在 Atlassian，工作就是开发 coding agents，所以大量试用 models。我从 GPT-3.5 开始尝试各种 models。那时让它编辑一个 file 都非常困难。 GPT-4 好一点，但仍然不到位。真正改变局面的是 Sonnet 3.5 v2。这是又一个 inflection point。过去三年里的几个转折点极大改变了我的 workflow。

A few of these inflection points really like changed my workflow a lot through the past three years. So, can you screen share and show us your current setup? Yeah. Yeah, sure. David, are you a terminal person or a GUI? I'm actually both. Like literally, there's there's weeks where I'm only using the terminal like CMAX. And there's weeks where I'm like either the Codex app or the cursor agents window. So, I'm a hybrid. Okay, cool, cool, cool. I'm more on the terminal side. 主持人：可以 screen share，给我们看看当前 setup 吗？ Kun：当然。 David，你更喜欢 terminal 还是 GUI？主持人：两者都会。有些星期只用 terminal，比如 cmux；有些星期主要用 Codex app 或 Cursor agents window，所以我是 hybrid。

So, I try to use the terminal as my primary uh development experience. I have just built so much muscle memory with it. Uh I have been using my setup for like over two decades writing codes in terminal. I I think the habit it creates for me is just to have my hands on the keyboards pretty much at all times. What would you tell to people who like are afraid of the terminal? You know, maybe there were not developers, they're not that technical. What would you say to those? Kun：很好。我更偏 terminal，把它当作主要 development experience，因为已经形成了大量 muscle memory。我在 terminal 里写 code 已超过二十年。它养成的习惯是让双手几乎始终放在 keyboard 上。主持人：对于害怕 terminal 的人，你会说什么？他们也许不是 developer，技术背景没那么强。

there is a learning curve initially to get used to a different kind of workflow uh where you are using your keyboard to control like all the movements and everything uh in the terminal. Kun：一开始确实有 learning curve，需要适应一种用 keyboard 控制 terminal 内所有移动和操作的不同 workflow。

So initially I think you will feel there is like a productivity loss uh because you are not comfortable and you are uh just you're not sure how to get everything done right so there is the initial phase but once you get past that it keeps you in the flow because you can just have your hands on the keyboard like control everything from here uh using keyboard shortcuts and everything your thoughts uh is like the only thing that's flowing awesome so walk us through it we have pi agent here what's happening the bottleneck isn't coding anymore. 最初你会… …感到 productivity loss，因为还不舒服，也不确定怎样完成所有操作。但跨过这个阶段后，它会让你保持 flow：双手一直留在 keyboard 上，用 shortcuts 控制一切，只有思路本身在流动。主持人：太好了，请带我们看看。这里有 Pi agent，发生了什么？旁白广告：现在 bottleneck 已经不是 coding，而是 review 数千行 AI 写出的 code，确保不会发布 slop。

The bottleneck is reviewing thousands of lines of AI written code to make sure you're not shipping slop. And a lot of you already use cloth code, cursor, or codeex to write code. But what you don't have yet is something that reviews your code as carefully as a human would. That is what code rabbit is. And code rabbit doesn't just take one look at your code through an LLM. Instead, you connect code rabbit to your repo and it understands your entire codebase. 很多人已经用 Claude Code、Cursor 或 Codex 写 code，但缺少像人一样仔细 review 的工具，这就是 CodeRabbit。 CodeRabbit 不只是让一个 LLM 看一次 code；连接 repo 后… …它会理解整个 codebase，提取 Linear issues 和 documentation 等外部 context，并运行 40 多种 linters 与 security scanners。

It pulls in external context like link issues and documentation. And on top of that, it runs 40 plus different llinters and security scanners. Then it surfaces it in one digestible format so you can act on it. And the feedback Code Radbit gives you isn't just add more tests. It's clear, specific, and actionable. It shows you what changed, why it matters, and how to fix it. 然后把结果汇总成可消化格式，方便采取行动。 CodeRabbit 不会只说“多加测试”，而会给出清晰、具体、可执行的 feedback，解释改了什么、为什么重要、怎样修复。发现 bug 时，还有 Fix with AI 按钮可一键解决。

And when it finds something like a bug, there's a easy fix with AI button that solves the issue with one click. And what's nice about Code Rabbit is that it meets you where you already work. Whether that is during PR review, in your IDE, in the CLI, in Slack or Discord, so you can review where is the most convenient for you. And it's not a static checklist either. Give it feedback once in plain English, and it remembers your standards and preferences every review after that. CodeRabbit 还能在你原本的工作场所提供帮助，无论是 PR review、IDE、CLI、Slack 还是 Discord。它也不是静态 checklist；只需用自然语言反馈一次，之后每次 review 都会记住你的 standards 和 preferences。

Even today, Code Rabbit is already reviewing over 3 million unique pull requests every single week, which makes it the most installed AI app on both GitHub and GitLab. And if you're building a public open source repository, Code Rabbit is free forever. All of us are shipping more code than ever before and it has never been more important to understand and review your pull requests properly. That is what code rabbit solves. 目前 CodeRabbit 每周已 review 超过 300 万个独立 pull requests，是 GitHub 和 GitLab 上安装量最高的 AI app。如果构建 public open source repository，CodeRabbit 永久免费。大家发布的 code 比过去任何时候都多，正确理解和 review pull request 也从未如此重要，这正是 CodeRabbit 解决的问题。

If you want to try code rabbit yourself, it's going to be the first link in the description. I just have a frameless terminal window here. Uh this is using uh western. I I like western because it's really highly customizable. I can just like change everything about it. Uh and here I made it frameless. So there is no window border, there is no nothing like it's just a pure terminal window and with like background blur and everything. I like it to um to look good. 想试用可点击 description 第一条链接。 Kun：我这里只有一个无边框 terminal window，用的是 WezTerm。我喜欢它，因为高度可定制，几乎什么都能改。我把它做成无边框：没有 window border，什么都没有，只有纯 terminal window，再加 background blur 等。

Um so when I work I I can look at a pleasing experience. Uh what I have here is a Herder session. Herder is kind of like a modern version of T-Max. I was using T-Max for like over 10 years and only recently discovered Herder and I just really like it. It's a more modern take how you manage multiple sessions in your terminal window. So have you tried like Cmax and what's the difference between like CMAX and Herder? Yeah, so Herder is even more in the terminal, right? 我希望它看起来舒服，让工作时拥有愉悦体验。这里运行的是 Herder session。 Herder 类似现代版 tmux。我用了十多年 tmux，最近才发现 Herder，非常喜欢。它以更现代方式… …管理 terminal window 里的多个 sessions。主持人：你用过 cmux 吗？它与 Herder 有何不同？ Kun：Herder 更彻底地位于 terminal 内。

CMAX has a GUI frame and uh within the frame there's like some PES that are terminals, right? But Herder is like entirely in the terminal. So the good thing about this is like this terminal window is just a program running, right? This is running on my Mac Mini and I can connect to this same session from my phone. Um, so I I can take my phone, I do a SSH connection, I can get this exact same terminal window. So everything stays the same. cmux 有 GUI frame，里面是一些 terminal panes；Herder 则完全处于 terminal。好处是，这个 terminal window 本质上只是一个运行中的 program。它运行在 Mac Mini 上，我也能连接到同一个 session。

With other GUI applications, it's harder, right? Um, because you can't quite just like remove desktop. Uh, and that's not going to be a good experience either. So this is like the main difference for me. I see. Real quick, if you want to implement Coon's entire setup for yourself, it's going to be available in the second link below the video completely for free. This includes his first mate setup herder westerm and everything else uses to actually build with AI agents. 我可以拿出手机，通过 SSH connection 获得完全相同的 terminal window，一切状态都不变。其他 GUI application 就困难得多，因为不太可能直接 remote desktop，而且体验也不好。这是对我来说最主要的差异。主持人：如果想完整复刻 Kun 的 setup，视频下方第二条链接免费提供，包括 FirstMate、Herder、WezTerm，以及他用 AI agents 构建产品的其他全部工具。

Again, grab it completely for free. Second link below the video. So here I have a herder uh session running and this in this herder session you can see on the top left there are spaces and I'll talk about that in a bit. It's basically like workspaces in her. You can use that to organize your work your different projects and uh etc etc. And on the bottom left there's agents. And that's a good thing about Herder uh which is that it understands what agents are. 再次说明，完全免费，链接在 description 第二项。 Kun：这里运行着一个 Herder session。左上方可以看到 spaces，稍后会讲，它们基本类似 Herder 中的 workspaces，用于组织不同 projects 等。左下方则是 agents。 Herder 的优点在于…

The previous like terminal multiplexers uh like T-Mox and uh Zelage they don't quite understand agents right they understand your terminal windows your tabs and uh PES they don't know what a session uh what agents are but Herder knows that I have a PI agent working here. The working status basically tells me that I don't need to like look at it. uh it's still working. It's not waiting on me. Uh so that's a very useful thing. …它理解 agents。以往的 terminal multiplexers，比如 tmux 和 Zellij，只理解 terminal windows、tabs 和 panes，并不知道 agent session 是什么。 Herder 则知道这里有一个 Pi agent 正在工作。 working status 告诉我无需查看它，因为它仍在运行，没有等待我。

Yeah, in this herder um I use herder to organize my agents and allow me to navigate across all the sessions I'm working on. Uh so that's the the thing I uh like about it. But most of the time I only work on one session which is this first mate. So this is a workflow that uh that was developed like maybe since a few months ago uh when I realized that I'm managing so many sessions all the time. 这非常有用。我用 Herder 组织 agents，并在所有工作 sessions 之间导航，这是我喜欢它的地方。但多数时间我只待在一个 session，也就是 FirstMate。这套 workflow 是几个月前开始形成的，当时我意识到自己一直在管理太多 sessions。

Similarly, I think uh when people use CMOX and other uh kind of orchestration uh apps, in order to get more work done, people have to manage multiple uh parallel agent sessions. Yeah. Right. Because one agent uh can take a while to get work done. Uh so people start to spin up more sessions and sometimes I see people have like 20 20 30 sessions. 同样，使用 cmux 等 orchestration apps 时，为完成更多工作，人们必须管理多个 parallel agent sessions。因为一个 agent 可能要很久才能完成，大家会开启更多 session，有人甚至同时开 20、30 个。

But then I kind of like uh was going crazy because yeah 20 sessions you have to keep that in your head right you need to remember oh what is this session what is it doing yeah so that pain caused me to develop first mate because I I I just don't think that's going to be the end game like I don't want to spend all day just juggling between the tabs right and remembering what was what I develop first mate because I think uh the agents as they become more capable should be able to juggle all those things for me. 但这快把我逼疯了。 20 个 session 都得记在脑中：这个 session 是什么，在做什么？这个痛点促使我开发 FirstMate。我不认为整天切 tabs、记住每个 tab 是什么会是最终形态。随着 agents 更 capable，它们应该能替我协调这些事情。我从几个月前开始试验这套 setup，结果真的可行。

And I started experimenting with this setup since a few months ago and it actually works. Uh so now this is like the only agent I talk to most of the time. Uh I just talk to first mates and let first mate manage all the other sessions for me. So I can maybe like walk through some examples here. Um so this is already hap something is already happening, right? It's doing some work. It's telling me no actions needed. 现在多数时间，我唯一直接交谈的 agent 就是 FirstMate，由它管理其他 sessions。我可以展示几个例子。这里已经有事情在进行，它告诉我“无需采取行动”。现在我想到一项之前开始的工作。我直接问：“Apple 是否已经批准这个 app 的 review？

Um and now I'm wondering about some some work that was uh happening a while ago. So I can just ask, hey, did Apple approve our app review for this app? So that's an iOS app uh I I was building and I submitted for app store uh review and um I'm not sure whether that's approved. So I can just check uh with first mate and first mate is not going to do that by itself. It's going to delegate the task to another um crewmate. ”那是我构建并提交 App Store review 的 iOS app，不确定是否批准，所以可以让 FirstMate 检查。 FirstMate 不会自己做，而是把 task 委派给另一个 crewmate。原因是如果 FirstMate 自己处理，就会变忙。

The reason is that if first mate does this for me, then first mate will get busy, right? And I cannot uh talk to first mate again. So now you can see uh captain I'm checking this now it's asked another crew mate to check it for me. The benefit is that I can talk to first mate now uh without it's like blocking on something else right I can just ask for something else. Uh I remember there are some good PRs uh sitting on me uh for review in treehouse. Can you check? 那样我就暂时无法继续与它交谈。现在可以看到它说：“Captain，我正在检查。 ”它已经让另一个 crewmate 去查。好处是我仍可继续与 FirstMate 对话，不会被其他任务 block。我还能问：“我记得 Treehouse 里有些不错的 PR 等我 review，可以检查一下吗？ ”

嘉宾 So first mate has access to all the projects you're working on. Yeah. Yeah. It's the coordinator of everything. Uh it knows about all my projects. I'll talk through some like scaling issues uh later on as well. 主持人：FirstMate 能访问你正在做的所有 projects？ Kun：对，它是所有工作的 coordinator，也知道我的全部 projects。

Uh the basically I have um like maybe 20 to 30 GitHub repos that are public and uh have quite quite some popularity and people file issues and PRs to me right if I am to manually juggle all the 20 30 projects I'll just go crazy um so I pretty much let first mate manage all the projects for me. Um so when I say treehouse uh first mate knows which project that is. My main question would be like how did you develop enough trust to delegate so much responsibility to first mate? 后面我也会谈到 scaling issues。我大概有 20 到 30 个 public GitHub repos，其中一些相当受欢迎，大家会提交 issues 和 PR。如果手动协调这 20、30 个 projects，我会疯掉。所以我基本让 FirstMate 管理所有 projects。当我说 Treehouse 时，它知道指的是哪个 project。

So trust uh takes time to build right. So I didn't initially just trust it for everything. 主持人：我的主要问题是，你怎样建立足够的 trust，把这么多 responsibility 委派给 FirstMate？

I started playing with it um initially as a prototype and I started like really letting it do more and more for me and I observe how it does and sometimes uh initially especially uh like when I just first began using it uh there were many problems uh such as like it's not delegating uh the right thing to the right crew mates or it's like sometimes having multiple crew mates doing the same thing conflicting which with each other things like that um and I started optimizing the process more and more and eventually got to a pretty sweet spot where I can just like let first mate handle all those for me and I don't see like mistakes or uh suboptimal routing anymore. Kun：trust 需要时间建立。我一开始并没有信任它处理所有事情，而是从 prototype 开始试。我逐渐让它做得更多，并观察表现。尤其刚开始时有很多问题：比如没有把正确 task 委派给正确 crewmate，或者让多个 crewmates 同时处理同一件事并彼此冲突。我不断优化 process，最终达到一个比较理想的状态，可以放心让 FirstMate 处理，而且不再看到明显 mistakes 或 suboptimal routing。

So now it's also uh started a task to uh to review some of the treehouse PRs, right? And now I can continue talking to it uh about some like anything else I want to do. Remember there was work happening around optimizing the workspace organization for first mate. Uh where are we? So I can just keep asking about these things uh that uh that are crossing my mind. 现在它又启动了一个 task，去 review 某些 Treehouse PR。我可以继续问任何其他事情。比如：“我记得之前有一项优化 FirstMate workspace organization 的工作，现在进展如何？ ”我可以不断问起脑中想到的跨项目事项。

The really good uh feeling I get um here right now like the thing I enjoy the most is that I'm constantly in a brain dump kind of mode. I have thoughts and I just dump my thoughts to first mates and let first mate do everything else. I don't need to worry about like all the steps it needs to take uh to get things done. And do you ever go into the specific like sub agents to look at what they're doing or not really? 这套方式给我最好的感受，是自己一直处于 brain dump mode：有任何想法就直接倒给 FirstMate，其余事情全由它处理。我不必担心完成任务所需的每个 step。主持人：你还会进入具体 sub-agents 查看它们在做什么吗？ Kun：现在极少。开发 FirstMate 的早期，我会经常进去观察。

Very rarely now. uh initially when I was developing first mate I did that a lot just to observe right uh is it doing the right thing uh is first mate uh communicating efficiently right uh did it actually bring the context to the crew mates uh so initially I did a lot of observation myself but now I don't do that much um but if I uh sometimes I will still feel like uh why is this taking so long right kind of like working in a real company in a real team sometimes the manager will still go to a direct report directly uh like maybe bypassing the tech lead just because uh sometimes it's like more efficient to directly communicate. 我会确认它是否做对、FirstMate 是否高效沟通、有没有把 context 传给 crewmates。早期我亲自观察很多，现在不太需要。但有时仍会觉得“为什么这么久”，就像真实 company 或 team 里，manager 偶尔会绕过 tech lead，直接找 direct report 沟通，因为这样更高效。如果想这样做，我可以在 Herder 中调出这个 view，它允许我跳到任何其他 agent session，看看里面实际在发生什么。

So if I want to do that, I can just like in her I can uh bring up this view. Uh this view pretty much like lets me jump to any other u agent session. Uh I can just jump to see like what is really happening there. Yeah, that's very very nice. Yeah. So here um basically uh first mate answered my last question. Uh this is at a strong design checkpoint. 主持人：非常好。 Kun：这里 FirstMate 已经回答了上一个问题：“当前处于一个很稳固的 design checkpoint。 ”我之前与 FirstMate 讨论过一个 idea，但没有继续 implementation，因为不确定这个 approach，总觉得哪里不对。

Okay. I discussed like an idea with first mates a while ago uh but we didn't proceed to implement uh because I wasn't sure about this approach. I felt like something was uh was off. Let me actually show you something here. So with this kind of technical design, sometimes it's a little bit complex, right? Uh it has lots of moving pieces. It has trade-offs. Uh has like pros and cons. 我来展示一下。 technical design 有时相当复杂，包含许多 moving pieces、trade-offs、pros 和 cons。这类情况我通常使用… … Lavish，也就是 interactive HTML artifact，帮助自己更容易理解 idea 和 trade-offs。

Um so this kind of case I usually use lavish uh a interactive um HTML artifact to uh allow me to more easily understand uh what the idea is and what the trade-offs are. Uh so let me uh try that now. Can you use lavish to review the um herder workspace design uh with me? What this does is that first mate will create a HTML artifact with visuals about uh the design decisions we have to make. Uh so it's going to be easier for us to collaborate on that. 现在来试一下：“可以用 Lavish 和我一起 review Herder workspace design 吗？ ”FirstMate 会创建一个包含 visuals 的 HTML artifact，展示我们必须做出的 design decisions。这样我们更容易协作。它通过 Lavish AXI 生成内容，可能需要一点时间。

Yeah, it's using the lavish uh axi to do this. It can take a little while uh to create the artifact. Right now you were typing but you know you were a big proponent of speaking. When do you do which? Uh I pretty much like always use uh voice input now. Um, the reason you saw me typing was I realized uh a while ago uh that if I use voice input, it actually interferes with my microphone somehow. I haven't fixed that yet. 主持人：你刚才在打字，但一直很支持使用语音。什么时候用哪一种？ Kun：现在我几乎总用 voice input。刚才打字是因为之前发现 voice input 会以某种方式干扰 microphone。这个问题还没修好。不过通常写 prompt 时我都用 voice input，唯一例外是需要 copy-paste URL、file path 之类的情况，把它们念出来没有意义。

Um, but mostly yeah, when I prompt I just uh use voice input. Um, the only exception I would say uh is when I have to copy paste a URL or a file path or something like that. Uh, it just doesn't make sense to speak that out loud. Yeah. Yeah. Of course. So now yeah it's writing the design review uh in a HTML and uh very soon we can probably like see that in my browser. So you're using uh 5.6 as your main agent. Why is that? Yeah. So 5.6 soul on x high as my first mate. 主持人：当然。 Kun：现在它正在把 design review 写成 HTML，很快大概就能在 browser 里看到。主持人：你用 5.6 作为 main agent，为什么？ Kun：我用 5.6 Sol、x-high 作为 FirstMate。

嘉宾 Uh so first mate uh the thing I uh learned is that first mate is kind of like juggling through a lot of context. It needs to rationalize a lot of different things like which thing am I talking about right? 我的经验是，FirstMate 需要协调大量 context，也必须推理很多不同事情：我在说哪件事？当我说 Herder workspace view design 时，它要知道这是之前处理过的那项工作，也要明白我指的就是它。

when I say herder workspace uh view uh design it needs to know oh a while ago we worked on that and that's what I mean um so first mate actually needs a lot of reasoning it needs to really uh be able to rationalize all the complexity and then I have different rules for different crew mates so when it's uh when first mate is dispatching a task to a crew mate it has a set of preferences I told it uh and it will use those ref uh those preferences to route the task here This file is basically where I write the rules and the rules will will basically tell first mate in what case should we use which agent uh which model and at what reasoning efforts. 所以 FirstMate 需要很强 reasoning。它必须能够理清全部 complexity。我还为不同 crewmates 设置了不同 rules。 FirstMate dispatch task 时，会使用我提供的一组 preferences 进行 routing。这个 file 就是我写 rules 的地方，它会告诉 FirstMate… …什么情况该使用哪个 agent、哪个 model，以及何种 reasoning effort。

Um it previously there was a um a rule here that says for any high complexity technical design and product design use uh fable uh in cloud code. Um I think fable has the depth and has the creativity that I really like. Uh so for those very complex designs I I like to use Fable as the crew mate. Uh but by default right now uh it's using uh 5.6. So purely because of the subscription practicality. So uh I can show you I have this thing tracking my uh quota. 之前这里有一条 rule：任何高复杂度 technical design 和 product design，都在 Claude Code 中使用 Fable。我喜欢 Fable 的 depth 与 creativity，所以这类复杂设计会把它当作 crewmate。不过目前默认使用 5.6。

I ran out of my gro uh kota uh and waiting for the reset. I'm almost out of the fable kota this week and the reset is long uh is still like a ways to go. So I'm saving this remaining fable kota for like things I really really need fable for. this whole um LLM quota from the subscriptions we get are just not enough. Um I think I would suggest uh the LLM companies to actually think about having a higher tier. Yeah. 500 because the $200 uh tier right now is not sufficient. 主持人：纯粹因为 subscription 的实际限制？ Kun：可以给你看，我有个工具追踪 quota。我的 Grok quota 已经用完，正在等 reset；这周的 Fable quota 也快耗尽，但距离 reset 还很久，所以要把剩余额度留给真正需要 Fable 的任务。 subscription 提供的 LLM quota 完全不够。

I agree completely. I mean people are going to hate the stake, but we do need some $500 tier, you know. Yeah. Yeah. So you never do API pricing. Yeah. So I I think for individuals it doesn't make sense to do API pricing. If I actually take API pricing for everything I worked on for the past month, it's going to cost more than like $10,000. It's just not sustainable. 我建议… … LLM companies 考虑推出更高 tier。主持人：500 美元？ Kun：现在 200 美元 tier 不够。主持人：完全同意。大家可能不喜欢这个说法，但确实需要某种 500 美元 tier。你从不用 API pricing 吗？ Kun：对于个人，我认为 API pricing 不合理。

I think another thing that might help uh you know like GPT 5.5 and 5.6 and a lot of the cloud models have a fast mode. Yeah. Right. Um the fast mode will basically cost more to get you lower latency. I want the opposite of that. 如果过去一个月的全部工作都按 API 计费，会超过 10,000 美元，无法持续。另一个可能有帮助的东西是：GPT-5.5、5.6 和许多 Claude models 都有 fast mode。

I want something that's slower but will be cheaper. uh because there are so many tasks that are like backgrounds tasks that I don't really care about it being finished very fast right I care about how much work can I get done in total because my my total quota is the bottleneck right now uh so I I wish uh they will uh eventually develop something like that yeah there is demand for this for sure different tasks require different tools and settings right like if you're working on something like I have this very important meeting in 10 minutes you know we have a transcript from last week you need the fast mode But like if you're just dispatching an agent overnight, you need the slowest of slow modes. 主持人：对。 Kun：fast mode 会支付更多费用换取更低 latency，我想要相反的选择：更慢但更便宜。因为很多 background tasks 根本不需要很快完成。我在乎的是总共能完成多少工作，因为 total quota 才是当前 bottleneck。希望最终会出现这样的产品。主持人：这肯定有需求。不同 task 需要不同 tools 和 settings。

You know, you don't care if it's five tokens per second. You just want it done. Yeah. Exactly. Uh I think that the whole spectrum uh ideally the options are available to us to make the trade-offs. So this is uh the question I asked earlier about the herder organization. Uh so this idea uh for some context uh is about how first mate uses herder to organize it uh crewmates. 比如十分钟后有重要 meeting，要立刻根据上周 transcript 做准备，就需要 fast mode；但如果 overnight dispatch agent，就需要最慢的 slow mode，哪怕每秒只有五个 token，只要最终做完即可。 Kun：完全正确。理想情况下，整个 spectrum 的 options 都应提供，让我们自己做 trade-off。

So right now everything is like a flat list and it's sometimes hard to understand like which agent is doing what. So there was an idea to do better organization. So here um basically it brought up this uh design review uh artifact. It has a recommendation. Um but let me uh walk through the actual uh proposals. Yeah, this is kind of the hierarchy we have today. 这里是刚才关于 Herder organization 的问题。背景是 FirstMate 如何使用 Herder 组织 crewmates。目前所有内容都是 flat list，有时难以理解哪个 agent 在做什么，所以有人提出更好的组织方式。

Uh it used uh this kind of like a tree view right to show me uh what this is. uh and this is very helpful otherwise uh what we usually do is like in the terminal we discuss with the agent and the agent will like just print a long wall of text right so here I can see this visually uh it's a tree structure uh this is um what it does today native now uh repository and uh work tree hierarchy uh yeah different work trees uh gets grouped into uh by by repo and there's a recommended uh architecture here uh so this is uh also So pretty uh cool. 它已经打开 design review artifact，并给出 recommendation。我来看看 proposals。这是当前 hierarchy，它用 tree view 展示，非常有帮助。否则平时在 terminal 里与 agent 讨论，它只会输出一整墙文字。这里可以直观看到 tree structure：当前 native hierarchy 是 repository 和 worktree，不同 worktrees 按 repo 分组；旁边还有 recommended architecture。

Lavish makes the agent actually produce a whiteboard. So this is a ecalraw whiteboard and I can pretty much like just look at uh all the uh diagrams. Um I can adjust this as well. I can like say this is not right. I I will uh move the boxes around uh and I can I can send the feedback back to the agent. Um so there's a lot we can do interactively uh on these artifacts and that's what I uh I really like about uh like using this lavish based uh workflow. 这很酷。 Lavish 让 agent 真正生成了一个 whiteboard。这是一个 Excalidraw whiteboard，我可以浏览全部 diagrams，也能自行调整。比如说“这里不对”，移动 boxes，再把 feedback 传给 agent。

Um so it at the end it will usually point out the decisions I have to make. Uh the open questions it has uh is uh ambiguous. Uh so it has a few approaches uh right I need to decide on one. I think I I want to prioritize the supervisor ownership uh because the grouping uh will make more sense uh when it's grouped by supervisor. There's a lot of context here, so I I I'll probably like not spend too much time on this. 这些 artifacts 可以进行很多 interactive 操作，我非常喜欢这种基于 Lavish 的 workflow。最后它通常会指出我必须做出的 decisions，以及仍然 ambiguous 的 open questions。这里有几个 approach，我要选一个。

Um, but I'll cue this decision and I'll send that back to the agent. Uh, and that's it. Like that's pretty much like the workflow to make a decision like this. So the main question, like you said, there's a lot of context, you know, with these different projects, you know them. How do you still achieve something like deep work? 我认为应该优先考虑 supervisor ownership，因为按 supervisor 分组会更合理。这里 context 很多，不打算花太久。我会把这个 decision 排进队列。再把决定发回给 agent。就是这样，这基本就是做这类 decision 的 workflow。

You know, how do you still go into the flow without, you know, going crazy with like even even with the setup, you know, you can still have too many faults and like, oh, I wonder how that project is doing. I wonder how what this agent is doing like how do you still achieve deep work? 主持人：你说这里有大量 context，也有不同 projects。你怎样仍能做到 deep work、进入 flow，而不被“那个 project 进展怎样”“这个 agent 在做什么”等过多思绪逼疯？

主持人 Most of the time uh when when you think about uh the time I was spending with first mates uh in this session it was mostly spent on ambiguous decisions. So I was not uh really like jumping between tabs or uh wondering hey like uh was that work still happening uh was that waiting on something. I don't need to worry about those things. The only thing I spend my time on is like truly ambiguous decisions. the things that uh really does need my judgment. Kun：回想刚才与 FirstMate 相处的时间，大部分都花在 ambiguous decisions 上。我没有在 tabs 之间来回跳，也不需要担心某项工作是否仍在进行、是否正在等我。唯一需要投入时间的，是真正 ambiguous、确实需要我 judgment 的 decisions。

So pretty much like all day what I'm doing is like I uh one is that I'm dumping my thoughts into first mate. Everything I want to do I just tell first mate I want to do that and then I I can forget about it. Uh it will either get done or it will come back to me as a decision to be made. So that's one thing I do just tell first mate all my thoughts. The other thing is like to deal with the decisions that I uh have to make that first mate cannot uh determine by itself. 我一天里大致做两类事。第一，把 thoughts 倒给 FirstMate。想做什么就告诉它，然后忘掉；事情要么被完成，要么作为需要决定的问题返回。第二，处理 FirstMate 无法自行判断、必须由我做出的 decisions。我给 FirstMate 写了很多 instructions，说明何时必须回来找我，因为这很难调好。

Uh I actually have a lot of instructions for first mates on when it should come back to me. Uh because that that is actually a very tricky thing to get right. Sometimes uh we see uh agents to just go wild um and they uh do um a lot of things autonomously without your really your eyes on it and at the end you will notice it did the wrong thing. 有时 agents 会失控，在你没有真正监督的情况下自主做很多事，最后才发现做错了。我们确实需要调整 agent behaviors，找出 sweet spot。而每个人对这个位置的 preference 可能略有不同，所以需要与 agent 充分磨合。

And uh I think that's one thing that we really need to tweak about our agent behaviors to figure out like what is the sweet spot and each person maybe have like maybe have a slightly different uh preference for where that is. Uh so that is something like we need to really uh talk through with our agent. So with this setup like how many tokens are you doing on a typical day? I actually didn't count the tokens. 主持人：用这套 setup，一天通常消耗多少 tokens？ Kun：我没数 tokens，主要看自己用了 quota 的百分之多少。现在会努力维持比较好的平衡。你看到我的 Claude quota 比预期更快耗尽，是因为 Anthropic 告诉我们 Fable 将被移除。

Uh I mostly count uh like how many percentage of uh the quota I'm using. Okay. Right now I try to keep a uh good balance. The reason you saw that uh I exhausted my cloth quota uh quick uh more quickly than I should um is that uh anthropic told us that label will be gone. Yeah, they they I I I even called them out. I said like what are these childish games like either keep it or remove it. 主持人：对。我甚至公开批评他们，说这些幼稚把戏算什么；要么保留，要么移除，不要每周延期。或者像 OpenAI 一样给我们可储存的 resets。这样就能自己控制。现在的状况很遗憾，Claude quota 快用完了。

Don't do this like weekly extensions, you know,or like give us the banked resets that OpenAI has, right? Yeah. So we can control it. Uh so unfortunate situation um almost out of cloud kota. There was like even like a couple people that went to the hospital. Have you seen that? Because they like stayed up for like two days straight because they thought Fable is getting removed and they ended up in the ICU. Yeah. Transparency. 甚至有人因为以为 Fable 要被移除，连续熬夜两天，最后进了 ICU。 Kun：我觉得 Anthropic 在 transparency 上确实很挣扎。每次与 developers 沟通重大变化时，都不够透明，也没给大家足够清晰的信息来做决定。这只是快速抱怨一下。平时我就靠这个留意 quota，确保消耗速度与 reset 节奏匹配。

I think this is something I see Anthropic really struggle with. Every time they communicate with developers uh like the way they have handled a lot of the major changes in the past, it's just not uh quite transparent enough and give everyone enough clarity to make their decisions. This is just like a quick rant. Uh, but usually I I just use this to keep my eyes on uh my quota. Uh, and I try to make sure I'm at a good pace with the resets, right? 使用 OpenAI 时可以更激进，因为我有几个 resets 可用，而且给得很多。主持人：但你没发现 GPT-5.6 Sol，尤其 extra-high 或 Ultra，在发布最初 24 到 48 小时里烧额度烧得非常快吗？

With OpenAI, I can be a little bit more aggressive. Uh, because I I have a few resets I can use. Uh, and they give so many resets. But didn't you find like GBD 5.6 Soul, especially on extra high or ultra, is like burning like crazy? At least it was in the first like day or 24 48 hours. So, there were a few things there. The first couple of days, OpenAI made a mistake. Um, they adjusted the context window, uh, the compaction threshold for 5.6 to 372K tokens, right? Kun：这里有几个原因。最初几天 OpenAI 犯了一个错误，把 5.6 的 context window compaction threshold 调到 372K tokens；后来才发现任何超过 272K tokens 的 request 都会被过度计费。主持人：哦。 Kun：所以很多人消耗 tokens 的速度都快于正常水平。他们后来为此做了 reset，也在修复。

And later they realized any request that goes beyond 272K tokens will get overcharged. Oh,so a lot of people are burning tokens more quickly than they should and they did a reset for that. Uh, so so they're I think they're fixing it. Uh but that was one of the reasons people felt uh 5.6 was burning tokens more um quickly than they uh used to. The other uh interesting thing was that altra is a special case. 我认为这是大家感觉 5.6 更烧 token 的原因之一。另一个有趣之处是，Ultra 属于特殊情况。 Ultra 和 Claude 里的 Ultra Code 未必只是 reasoning effort level。

Yeah. So I think altra and the ultra code in claude right they are not necessarily a reasoning effort level. They are a prompt that tells the agent to use sub agents aggressively, right? To to fan out tasks and uh for a while I think maybe even right now open AAI uh codeex when it uses ultra it will spin up so many sub agents and every sub agent is an ultra agent. It's just burning tokens way too fast. Uh I think X high and high are very different. 它们更像是一段 prompt，要求 agent 积极使用 sub-agents，把 task fan out。至少有一段时间，甚至现在，Codex 使用 Ultra 时会启动大量 sub-agents，而且每个 sub-agent 也都是 Ultra agent，烧 tokens 的速度太快。 X-high 与 high 很不同，X-high 并没有那么耗我的 token。

主持人 X high does not burn my tokens that fast. X high is actually uh my sweet spot because despite it being X high, it's actually pretty fast. Okay. Right. Uh the prompts that we just went through, it usually comes back pretty quickly. Oh, actually yesterday on that point I did testing with like some people and I was observing like they were running 5.6 Luna and I was like why is it so slow they switched to 5.6 Soul medium and it was faster. X-high 其实是我的 sweet spot。尽管叫 x-high，它反而很快；刚才那些 prompts 通常都迅速返回。主持人：昨天我和一些人测试时，看到他们运行 5.6 Luna，觉得怎么这么慢。切换到 5.6 Sol medium 后反而更快。

So I don't know what's happening at OpenAI but literally Luna is slower than Soul right now. 我不知道 OpenAI 在做什么，但 Luna 现在确实比 Sol 慢。

Yeah. So you know uh why that's the case there's a uh benchmark I can show here it's called deepu um I think you probably have seen this as well right uh deepu this is like probably like the benchmark I currently trust the most okay um because it's not contaminated yet it's new enough that the training data does not have this um so maybe eventually it will need a refresh but right now I trust this a lot and you can see um there are a few really interesting outliers here uh clock solid five. Kun：我知道原因。有一个 benchmark 叫 DeepU，你大概也看过。这可能是我目前最信任的 benchmark，因为它还没有 contamination，足够新，training data 里不存在。未来也许需要更新，但现在我很信任它。这里有几个很有意思的 outliers。

If you look at sonnet five up until like from low, medium, high, x high, it's like it's a reasonable curve, right? Um so uh the uh this graph is like the the left hand side is more expensive and uh the top end is more intelligent. Uh so the the top right is good. So set five is reasonable up until f high. But if you do max you see It's going crazy. It's uh it's the most expensive model there is. Solid 5 is more expensive than Fable. 看 Sonnet 5，从 low、medium、high 到 x-high，曲线都还合理。这张图横轴越往左越昂贵，纵轴越高越 intelligent，因此右上角最好。 Sonnet 5 到 x-high 都合理，但使用 max 后就失控了，变成最昂贵的 model；Sonnet 5 甚至比 Fable 更贵。

Yeah. So I think the reason this is happening is that uh when the model itself is not intelligent enough, it will just waste a lot of cycles, right? Doing the wrong thing. Uh especially when you ask it do to do max reasoning, it will think very hard, but it's still not intelligent enough to solve the problem. So it will just waste a lot of time and tokens. Um so I think it's the same with Luna. The unique thing about Luna is that it's a really tall uh curve. 我认为原因是，当 model 本身不够 intelligent 时，它会浪费大量 cycles，做错误的事。尤其要求 max reasoning 时，它会非常努力地思考，却仍不够 intelligent，无法解决问题，于是只会浪费大量时间和 tokens。我认为 Luna 也类似。

It means it stretches between um like very very low intelligence to very high intelligence. Luna 的独特之处是曲线很高，也就是从非常低 intelligence 拉伸到非常高 intelligence。

So I think there's probably a similar thing happening with Luna where if you ask Luna to work on a hard problem, it will just like not be able to solve it but it will waste a lot of token trying to. Yeah. Yeah. My observation is exactly that you you did like way too many tool calls to achieve something relatively simple and then you know 5.6 soul on medium just did it in like a minute. Exactly. Uh so right now I pretty much only use soul for my day-to-day tasks. 所以让 Luna 解决 hard problem 时，它可能无法解决，却会浪费大量 token 尝试。主持人：我的观察完全一样。它为相对简单的事情做了太多 tool calls，而 5.6 Sol medium 一分钟就完成。Kun：没错。现在 day-to-day tasks 我几乎只用 Sol。

Um, so I can like adjust the reasoning level for soul, right? Um, I find very little reason to use terra because uh why do I use terra when I can just adjust the reasoning level down for soul then I get more intelligent model uh at a lower cost. I use Luna uh in some cases though uh in my home assistant uh so I have a home assistant running my house and I can control my lights and uh everything right music players etc etc. uh that I use Luna because it's fast. 我可以调整 Sol 的 reasoning level。几乎没有理由用 Terra，因为把 Sol 的 reasoning level 调低，就能以更低成本获得更 intelligent 的 model。不过我在 home assistant 的某些情况会用 Luna。

I need it to be really really fast. Why why not like a open source model for that? Like you know Kim K2.7 code with like the naturals to fix on open router you can get like 200 300 tokens per second. Yeah. Yeah. So open source models um so there is a few ways to uh to do that. One is to run that locally. Actually I think uh for this kind of home assistant scenario I can actually run something locally on my Mac Mini and get it to work reasonably well. 我用 home assistant 控制家里的 lights、music players 等，所以需要 Luna 的速度。主持人：为什么不用 open source model，比如 Kimi K2.7 Code，通过 OpenRouter 上的 provider 能达到每秒 200、300 tokens？ Kun：open source models 有几种使用方式。

Right. Um the thing is that my Mac Mini right now is a very precious resource. Um because I do everything else on my Mac Mini, I I I have a lot of tasks to be done, right? Uh so it's uh I when I build the iOS apps, when I build everything else I'm doing, uh it consumes the compute from my Mac Mini and I I connect my phone to it. I connect everything else to it. Mac Mini is kind of my personal like uh compute. 一种是在本地运行。 home assistant 这种 scenario，确实可以在 Mac Mini 本地运行并取得不错效果。问题是，Mac Mini 现在是一项非常珍贵的 resource，因为其他一切都在那里做。我有大量 tasks，要构建 iOS apps 和其他所有东西，都会消耗 Mac Mini 的 compute；手机和其他设备也都连接到它。

Makes sense. Um and uh if I run a uh open source model uh on that uh it will just like compete with everything else I'm doing. Uh so that's one constraint. Um I could also run open source models through um cloud providers, right? Um but I find that costsaving not uh meaningful enough for me to really like switch away from something like Luna. Yeah. Uh so so so far I haven't really um like uh got into um local LMS and uh open source models that much. Mac Mini 就像我的个人 compute。如果在上面运行 open source model，就会与其他工作竞争资源，这是一个 constraint。我也可以通过 cloud provider 使用 open source models，但发现 cost saving 没有显著到足以让我离开 Luna。

And in terms of the harness, is PI your favorite harness right now? Yeah, so um there's a few harnesses I use right now. I'm using PI because uh I'm mostly working with GPT 5.6 today uh because the quota. But uh what I do right now is that if I use anthropics models, I use cloud code because they ban everything else. If I use GPT 5.6, I use that in pi. Uh the reason is that the codec COI is not that good uh as a harness. Uh it has some strength. 所以目前还没有深入 local LLMs 或 open source models。主持人：harness 方面，Pi 是你目前最喜欢的吗？ Kun：我现在用几个 harness。因为 quota 原因，今天主要使用 GPT-5.6，所以在用 Pi。

Uh for example, it's uh got out of the box integration with the image generation uh which is very handy. A lot of the other capabilities uh such as like managing background processes uh it's just not as good. It doesn't have all the bells and whistles uh that make the COI experience really smooth. If I compare that with pi, pi is highly customizable, right? So everything I want to achieve, I can pretty much just build a plug-in to achieve that. 如果使用 Anthropic models，我会用 Claude Code，因为他们封禁了其他 harness；如果用 GPT-5.6，就放在 Pi 里。原因是 Codex CLI 作为 harness 没那么好。它有一些优势，比如开箱即用集成 image generation，非常方便。

Um, so that's the main reason I choose pi uh over codeex for GPT models. Um, for Grock, uh, is the other model I use because I have a X subscription already and there's some quota there. I actually find Grock, uh, 4.5 really good. Yeah, I mean I'm in the EU, so you know I cannot use it because it's banned in the EU, but Oh, okay. That's unfortunate. Uh but uh yeah, so um it's a pretty good model. The way I think about Grock 4.5 is that it's Opus but on fast mode. 但管理 background processes 等许多 capabilities 不够好，也没有让 CLI experience 真正顺滑的完整功能。相比之下，Pi 高度 customizable，想实现什么都可以构建 plugin。因此使用 GPT models 时，我选择 Pi 而不是 Codex。另一个会用的 model 是 Grok，因为我已经订阅 X，那里有一些 quota。

Damn. Um yeah, it's pretty much capable as doing everything opus can do. Uh and it's super fast. And uh the interesting thing about using Grock, especially in the Grock harness, the Grock build harness is that uh it seems Grock build gives you free X API access to like read post and search for things. Okay,which otherwise would cost you um because X API cost money. So that's another benefit. 我觉得 Grok 4.5 非常好。主持人：我在 EU，不能用，因为那里禁止了。 Kun：很遗憾。它确实是个优秀 model。我把 Grok 4.5 理解为 fast mode 的 Opus。主持人：厉害。 Kun：它几乎能做 Opus 会做的一切，而且速度极快。

I I can use Grock build to like search for tweets and rationalize like news uh things like that for me. I think it's like one of the fastest improving harnesses as well, right? Yeah. Yeah. Totally. It's very young, right? It's it's very new and uh they only recently started being really serious about that and it's already got like better capability than codeex at some of the uh background process polling uh etc. I see. Okay. Any other Asian harness? 在 Grok harness，特别是 Grok Build 中使用还有一个有趣点：它似乎免费提供 X API access，可以读取 posts 和搜索。主持人：好。 Kun：否则 X API 本来要付费。所以我可以用 Grok Build 搜 tweets，分析 news 等。主持人：它也是进步最快的 harness 之一吧？

Uh I also use open code quite a bit. So open code is also pretty smooth. Uh I like the TUI experience there. Um but uh I think over time I started leaning towards Pi for more and more things. Uh because Pi uh is like more deeply customizable. Uh you can pretty much change every aspect about Pi and uh I kind of like about that. So like you are big on customizability, right? Kun：完全同意。它非常年轻，最近才开始认真投入，却已经在 background process polling 等方面比 Codex 更强。主持人：还有其他 agent harness 吗？ Kun：我也常用 OpenCode，TUI experience 很顺滑。

So do you think in the future more and more people will have their own setup instead of like these predefined you know packaged apps? I think likely uh there's going to be a spectrum of preference. uh what I uh think will happen is that there will be a group of people who are not very opinionated about uh how things should work. Uh they are kind of looking for others like give them proven playbooks. So they just want something that can work out of the box. 不过随时间推移，越来越多事情会倾向 Pi，因为它可定制得更深，几乎每个方面都能修改，我很喜欢。主持人：你非常看重 customizability。未来会不会有更多人拥有自己的 setup，而不是使用预定义、打包好的 apps？ Kun：我认为会出现一个 preference spectrum。

They don't have to worry about tinkering with that all day. Um that I think will always exist and that's that's good and uh people will build tools and publish that to those people uh so they can just use something that can work out of box. There will be another group of people who will start to have more and more personal preferences for how things should work and those people will need uh things that can really uh be customized and evolve. 有一群人对工具如何工作没有强烈意见，希望别人提供经过验证的 playbook，只想要开箱即用，不愿整天 tinkering。这类需求永远存在，而且很好；会有人为他们构建和发布 out-of-box tools。另一群人会逐渐形成更多 personal preferences，需要能够真正 customization 并持续 evolution 的东西。

One thing I showed you earlier was this quota widget I have on the menu bar, right? Yeah. This thing I I call it baby menu. I I built this myself and this thing is self modifiable. So right now you can see it's a bunch of things that would never exist together in a product. There's my GitHub uh stats right and there's like CPU and memory and then there's like my cloud and codeex quota. No one will build a product like this and ship it to other people. 刚才展示的 menu bar quota widget 就是例子，我叫它 Baby Menu，是自己构建的，而且它可以 self-modify。现在它把一堆绝不会同时出现在普通 product 里的东西放在一起：GitHub stats、CPU、memory、Claude 和 Codex quota。

Um this is this belongs to me and the reason uh I I have it like this. Uh the way I got it to become like this is through talking to baby menu. So I uh baby menu started as empty uh thing and I tell baby menu what I want and baby menu will self modify and uh start to become uh like that. I that's I think one version of the future that I think will happen to a lot of software. They will ship a reasonable experience out of the box. But now everyone has agents. 没人会构建这种 product 再卖给别人；它只属于我。它之所以变成这样… …是因为我直接与 Baby Menu 对话。它最初是空的，我告诉它自己想要什么，它会 self-modify，逐渐变成现在这样。我认为这是未来许多 software 的一种形态：先提供合理的 out-of-box experience，但既然每个人都有 agents，就应该能与 agents 对话并定制自己使用的软件。

So everyone should be able to talk to their agents and customize the software they use. And also like a lot of self-healing loops, you know, like a lot of software will have like a way to easily report a bug that starts like a codex CLI, implements it, goes into PR, you have like something like code rabbit, review it, and the bug could be fixed within 20 minutes without any human involvement. Yeah. Yeah. Totally, totally. 主持人：还会出现很多 self-healing loops。 software 可以方便地 report bug，自动启动 Codex CLI、实现修复、提交 PR，再由 CodeRabbit review；20 分钟内无需人类参与就修好 bug。

主持人 Uh something else that's very interesting that's happening uh is like first mate. First mate is basically uh a agents.md file. So there's the agents.mmd that prescribes how first mate should behave, how it should coordinate different tasks across a larger number of crew mates. So this is like the instruction for first mate. It also has a bunch of script files. So in this uh binary folder uh it has a bunch of uh bash scripts. Kun：完全同意。另一个很有意思的变化也发生在 FirstMate。 FirstMate 基本就是一个 AGENTS.md file。 AGENTS.md 规定 FirstMate 应怎样表现、怎样在大量 crewmates 之间协调 tasks，这就是 FirstMate 的 instruction。

These bash scripts basically handled a lot of the deterministic steps that first mate would otherwise have to take. um so bundle them into a bash script so that uh first mate does not need to spend tokens going through all those deterministic steps. Now the good thing about this is that both the agents.md and this bash scripts are in the current directory. So first mate can modify this right. 它还有一组 script files；bin folder 中有很多 Bash scripts，负责处理 FirstMate 原本必须执行的 deterministic steps。把这些步骤打包成 Bash script，FirstMate 就不必花 tokens 逐步完成。

Um, so one interesting thing I didn't even realize was going to happen was that when first mate has a bug um that's like preventing it from doing something uh correctly, right? It will just work around the bug by itself. So now this software is pretty much like unstoppable. There's no way you can stop first mate from doing what it needs to do. 好处是 AGENTS.md 与 scripts 都位于 current directory，FirstMate 自己可以修改。一个我甚至没预料到的现象是，当 FirstMate 遇到妨碍正确执行的 bug 时… …它会自行绕过 bug。现在这个 software 基本 unstoppable。

Even if this scripts have like various kind of bugs, it will maybe make first mate a little bit more inefficient, it will not stop it from working. This is a really big change from traditional software, right? Uh in a traditional software, if you have a bug, then it's a bug. It may be like breaking the app, it may be crashing the app, you can never get through it. 即使 scripts 有各种 bugs，也最多让 FirstMate 低效一点，不会阻止它工作。这与 traditional software 相比是巨大变化。

But now we have this like intelligent software that's pretty much uh always flexible and it's very hard to write a bug that can stop it from working. So basically then the challenge becomes like where are all the things that we humans are slowing down the agents and how can we remove ourselves from as many of them as possible only being involved in the ones that really matter. Yeah. Yeah. Totally. 传统 software 有 bug 就是 bug，可能破坏或 crash app，让你无法继续；现在出现了 intelligent software，它始终保持 flexibility，很难写出一个能彻底阻止它工作的 bug。主持人：所以挑战变成：在哪些地方是 humans 在拖慢 agents？怎样尽可能移除自己，只参与真正重要的环节？ Kun：完全正确。

嘉宾 What else about your setup should people know? Um I think maybe the other thing is uh no mistakes. Um so what we very very often I run into this so much is that I get AI to generate code for me right? Uh and uh Fable and uh 5.6 six they can write code very very quickly and very well they can do very complex changes but now if there is a complex change uh that's made by AI how do I know it's okay to merge it right 主持人：关于你的 setup，还有什么需要知道？Kun：另一个是 No Mistakes。我经常遇到这种情况：让 AI 生成 code，Fable 与 5.6 都能又快又好地写，完成非常复杂的 changes。但当复杂 change 由 AI 完成后，我怎样知道它可以 merge？

主持人 um so let me show you a very uh real example here I'll go to my other project called high bit this is a uh a AI harness I'm building for uh children uh to learn to use agents I just did a whole bunch of changes you can see here this is the diff I made a bunch of changes using uh fable uh yesterday and now I have this uh change sitting here. What do I do now? Do I go review every uh every line of this diff? It's going to take a long time, right? 我展示一个真实例子。这个 project 叫 Highbit，是我为 children 学习使用 agents 构建的 AI harness。昨天用 Fable 做了大量 changes，现在 diff 放在这里。接下来怎么办？我要 review diff 的每一行吗？那会花很久，而时间非常有限。

And our time is like really limited. So if I um review AI generated code, then there is an upper bound of how much work I can get done, right? Because uh it pretty much depends on how much code I can review every day without going insane. 如果由我 review AI-generated code，能完成的工作就有一个 upper bound，因为上限取决于自己每天能 review 多少 code 而不发疯。所以现在，我会把所有这些 changes 发给 No Mistakes。

So now what I do is that I have pretty much I send all these changes to no mistakes which performs a pipeline uh where it does adversarial review and validation and everything and present me with something uh that I can more easily determine whether it's okay. Uh so the way I do that is just uh nm. 它运行一个包含 adversarial review 和 validation 等步骤的 pipeline，再把结果以更容易判断能否接受的形式交给我。

主持人 Uh this is a uh shortand for no mistakes uh dashy which pretty much like just uh means get all the code changes that are in the current working directory into a branch and pass that to no mistakes. Uh so I just run that to show you. 使用方式就是运行 `nm -y`；`nm` 是 No Mistakes 的 shorthand，`-y` 表示把 current working directory 中的全部 code changes 放入 branch。

No mistakes will create a branch will ask the agent to create a branch for me and create a commit and uh like put all all the necessary information in there and then push that branch into a local git proxy where it performs the validation in a isolated work tree. Um so we can see that in a bit. Uh yeah it's uh already created um the uh commit. 然后交给 No Mistakes。它会要求 agent 创建 branch、commit，放入所需全部信息，再把 branch push 到 local Git proxy，在 isolated worktree 中执行 validation。

So this is the no mistakes pipeline. it first uh will analyze the intent and the way it uh understands the intent uh is by analyzing the agent session that produced this change. So the agent session is where I worked with uh fable uh to create this code change right I uh sent my prompt I expressed my intent originally there. So the first step is to understand that intent because that intent is the true requirements that should be respected. 这里已经创建了 commit。 No Mistakes pipeline 首先分析 intent。它通过分析产出这些 changes 的 agent session 来理解 intent。那是我与 Fable 合作创建 code change 的 session，original prompt 在那里表达了我的 intent。

It's what I said I want right it understands the intent. It does a rebase uh on the the the latest uh remote origin uh main branch uh so we don't run into merge conflicts later on. And now it's doing the adversarial review. So here it's using GPT 5.6 soul on medium to review the change. If it found something it will either autofix if it's a very obvious bug or it will escalate to me for approval if it's not obvious enough. If it has some product implications. 第一步必须理解它，因为 intent 才是真正应被尊重的 requirements。理解 intent 后，它会基于最新 remote origin main branch 做 rebase，避免之后产生 merge conflicts。现在正在运行 adversarial review，使用 GPT-5.6 Sol medium 检查 change。

Sometimes we find a bug and the solution to fix the bug will require changing the product, right? Those are the cases I want to actually have a say. I set some rules to say for those cases escalate to me. And uh here we can maybe see uh whether it will just autocomplete or escalate something. 发现问题时，如果是明显 bug 就 autofix；如果不够明确… …或带有 product implications，就 escalate 给我 approval。

Uh but basically this review step will keep going for a while and it will pretty much catch all the edge cases and bugs and missing uh things and uh at the end after the adversary review it will do testing as well. uh and the testing uh will produce some visible evidence that can really tell me it's really working uh as intended and then documentation linting and will push for PR uh and will babysit the CI pipeline until the PR screen and then it will tell me okay now you can go look at that um so usually when I launch no mistakes I don't look at this view I I just go away and do something else now nowadays I don't even launch no mistakes myself I ask first mate to do Uh so it's first mates launching this for me. 有时发现 bug 后，修复方案要求改变 product；这类情况我希望拥有发言权，所以设置了 rules，要求它 escalate。这个 review step 会持续一段时间。它会寻找 edge cases、bugs 和遗漏项。 adversarial review 后还会 testing，产出 visible evidence，真正告诉我是否按 intended 方式工作；随后运行 documentation、linting，push PR，并 babysit CI pipeline，直到 PR… …变为 green，再告诉我可以查看。通常启动 No Mistakes 后，我不会看这个 view，而是去做其他事。

Yeah. But pretty much this is like something that I think helped me a ton. Um because otherwise uh I would have seeing all the problems no mistakes cut told me that uh my code bases would absolutely go into a mess uh if I didn't use something like this. But doesn't it on large project like burn up too much limits too many tokens? 如今甚至不再亲自启动，而是让 FirstMate 启动。这个工具对我帮助巨大；否则从 No Mistakes 发现的全部问题来看… …如果没有类似工具，我的 codebases 绝对会变得一团糟。

Because for example, this reminded me of DeepSack from Versell and they have like so many warnings that like if you run this on a large code base it's going to burn tens of you know tens of thousands of dollars of API credits because of like how deep it goes. So how do you like balance the number of review it does versus you know burning your usage? Yeah. Yeah. Good question. Right. So I I think there is like there is one aspect which is the inevitable cost of quality. 主持人：但在 large project 上，它不会耗掉太多 limits 和 tokens 吗？这让我想到 Vercel 的 DeepSeek，有许多 warnings 说，在 large codebase 上运行会消耗数万美元 API credits，因为分析非常深入。你怎样平衡 review 数量与 usage？ Kun：好问题。这里有一部分是质量不可避免的成本。

Um so think of a software team. If we remove the code review process between humans, everyone will be able to merge code faster, right? But then you will have all kinds of like quality problems uh incidents that will happen in production sites and then eventually you have to pay the cost in another way. 设想一个 software team，如果移除 humans 之间的 code review，所有人都能更快 merge，但会出现各种 quality problems 和 production incidents，最终还是要用另一种方式付出成本。

嘉宾 Yeah. So that's the inevitable part. I think uh we have to make a trade-off. How much do we care about the quality of a certain project? 所以必须做 trade-off：某个 project 的质量有多重要？

Um and in some projects for example it's a demo for a weekend uh showcase it doesn't really matter that much right in those cases we can probably don't like skip running this kind of heavy validation uh pipeline but if it's building a production software and there are a lot of users that would be affected uh there are customers who will stop paying if we like break their experience that's the case we want to be careful um so I think there's a trade-off to be made on which projects to apply this kind of approach The other thing I uh I do is that I don't send every single change into this pipeline. 如果只是周末展示的 demo，也许不太重要，可以跳过这类 heavy validation pipeline；但如果构建 production software… …有大量 users 会受影响，破坏体验会让 customers 停止付费，那就必须谨慎。因此要根据 project 决定是否采用这类 approach。我也不会把每个 change 都送进 pipeline。有时只是 very simple fix。

Uh sometimes I make a very simple fix, right? Yeah. I'm confident enough that it's not going to break and I can make a judgment call to say this change let's just merge. Um so there are cases like that as well. But yeah, like the vast majority of my changes will go into this pipeline. I can show you my stats here. I I have this uh no mistakes stats. 如果很有信心不会破坏，就可以自行判断直接 merge。不过绝大多数 changes 仍会进入 pipeline。我这里有 No Mistakes stats，记录所有提交给它的 changes，并用一组 counters 理解表现。

I keep track of all the changes I sent to no mistakes and I have like a bunch of counters to help me uh understand how my changes are doing. That's another great thing about AI is like you can measure like so many things that like otherwise would be a hassle to measure thanks to like agentic engineering. It just you know you can start tracking literally everything about your work. Yeah, that's a really good point. Yeah. So with humans it's harder to track. 主持人：这也是 AI 的一个优点，可以衡量许多过去统计起来很麻烦的东西。借助 agentic engineering，可以开始追踪工作中的几乎一切。 Kun：对，这是很好的观点。 human 很难追踪；如果有 100 个人替我工作，想了解每个人做的全部事情会很困难，除非像 Mark Zuckerberg 一样记录一切。但 agents 的一切都能自动化，也都可以计数。

Like if I have a hundred humans working for me, I want to know everything they do. Uh it's going to be hard, right? Unless I'm like uh Mark Zuckerberg who like records everything. Um but uh yeah so with agents it's just like everything can be automated. We can count everything. That's really awesome. So this is my no mistake stats. This is like maybe the past three months or something like that. 这是我的 No Mistakes stats。这是过去大约三个月的数据：59 个 repos、总计 1,000 次 changes，其中 63% 被 No Mistakes 发现问题。主持人：你猜哪个 process 最有效？

Uh total changes a thousand across 59 repos. 63% of the changes got a mistake cut. And okay why do you think that is like if you had to guess which part of the process is it like a adversary review by a different model which part is the most effective? Yeah so this is this this by yeah fixes by step the review step caught and fixed most problems. So what I noticed is that um I think ever starting from GPT 5.5 the GPT model got really good at catching edge cases. 是另一个 model 做 adversarial review 吗？ Kun：这里按 step 统计 fixes，可以看到 review step 发现并修复了最多问题。我注意到从 GPT-5.5 开始，GPT model 变得很擅长发现 edge cases。

I compared a few models and I I like GPT the most uh as a reviewer. Uh it's really thorough. is really good at identifying edge cases that uh that can happen in in like a rare scenario but will happen. So those kind of u adversarial review from GPT uh just help a lot in catching these kind of problems. Another big hitter is documentation. 我比较过几个 models，最喜欢 GPT 作为 reviewer；它非常 thorough，善于识别只在 rare scenario 中出现、但真实会发生的 edge cases。 GPT 的 adversarial review 对发现此类问题帮助很大。另一个高产步骤是 documentation。

Uh so very often uh what I notice is that I will make a change the agents will make a change and it will not update for example the readme file which is still saying something that's like stale that's conflicting with the change we made right so those cases all got caught by uh this step which is really helpful very nice I mean one thing that comes to minds when I see this like advanced setups is that I know so many people from my audience who just spend all their time on their setup right like they literally have like this crazy agentic engineering setup but they never ship anything. 经常是 agents 做了 change，却没有更新 README，里面仍写着 stale、与新 change 冲突的信息。这些都能在 documentation step 被发现，非常有用。主持人：看到这种 advanced setup，我会想到许多观众把所有时间都花在 setup 上。他们有非常复杂的 agentic engineering setup，却从不发布任何东西。

So like how do you keep yourself like in a rational way where like okay maybe 10 20% of your time is spent on improving your setup and building these scripts and you know systems and other stuff is like you're actually building products with it. Yeah. Yeah. So that's a very very good topic. Um because uh I I I I see even myself a few months ago I was struggling with that. Everything takes time and when I I'm building something I run into some friction. 你怎样保持理性？比如只用 10%–20% 时间改进 setup、构建 scripts 和 systems，其余时间真正做 products。 Kun：这是非常好的话题。几个月前我自己也为此挣扎。一切都需要时间。构建东西遇到 friction 时，我会想做 tooling 消除 friction；但 tooling 本身也要花时间，还会产生需要修复的问题。

I want to build the tooling to remove that friction but building the tooling will take time from me right and then the tooling will have problems. I have to fix those problems. And then for a while for a little while I think a week or two I found that I'm mostly just working on tooling. I'm not building the real things. Um, so that changed after I started using first mate. 有一两周，我发现自己几乎只在做 tooling，没有构建真正的东西。开始使用 FirstMate 后情况改变了。 bug fix、遇到问题需要改进等 mundane things，我几乎不用花时间，只需告诉 FirstMate：“这里有个问题”，它就会修好。

So the good thing about first mate is that the mundane things like bug fix or like um I run into this problem and I just need that to be improved. I don't really need to spend much time at all. I just say to first mate, hey there's a problem and first mate will get that fixed. the things that take my time will naturally gravitate towards the more interesting and ambiguous decisions which tends to be new product development. 需要我投入时间的内容会自然转向更有趣、更 ambiguous 的 decisions，通常就是 new product development。现在 FirstMate 替我管理全部 mundane things，我就有更多时间构建真正想做的 products，比如 Highbit 和 iOS app。

So now with first mate managing all the like mundane things for me I have more time to actually work on the products I want to work on uh like high bits and my uh iOS app a thing I did not intend but it was emergent after I started doing this with first mate. 这不是原先计划的结果，而是在使用 FirstMate 后 emergent 出来的。主持人：这也是我喜欢你 setup 的地方。看到这些东西都源自真实 problems，令人耳目一新。

Yeah, I think that's like what I love about your setup is that it's also refreshing to see like a lot of these things actually emerge from real problems because there's a large percentage of people in the AI space which just like do things for the sake of it, you know, it's like I connected this together or like I built this setup, look how it works, you know, look how many agents it can run, but like it's not from first principles. AI 领域有相当一部分人只是为了做而做。他们会说“我把这些连接起来了”“看看这套 setup 能运行多少 agents”，但不是出于 first principles，也不是因为完成工作真正需要，只是觉得并行跑 200 个 agents 很酷。

It's not because they needed to get their work done. It's just like oh wouldn't it be cool if these 200 agents were running in parallel? But like as you're going through it, I can see reasoning for everything. Yeah. Yeah. So that that's definitely like my approach. I initially I didn't even want to build any of these tools. 你的每个选择则都有理由。 Kun：这确实是我的 approach。起初我根本不想构建这些 tools。三个月前，我刚辞掉 big tech 工作。我准备 solo 构建几个有潜力的 B2C apps，比如给 children 使用的 AI tutor。

I very recently uh three months ago I quit my um big tech job uh to go solo and my intention was to build a few uh B2C apps that I think have potential uh like the AI tutor for uh children and things like that. I have a few ideas I really want to build and that was my plan uh to spend all my time building those. The interesting thing about building solo versus working in a big company is that in a big company there's a lot of other bottlenecks slowing you down. 原计划把全部时间花在这些 ideas 上。 solo build 与 big company 工作有个有趣差别：大公司存在许多其他 bottlenecks。你会把时间花在 meetings… …以及与别人协调。我之前升到相当 senior 的 level，结果全部时间都用来告诉别人做什么，而不是自己动手。

Um so you spend your time uh in meetings coordinating with other people, right? Uh and I got to a pretty senior level and uh what happened was that all my time is spent on telling other people what to do instead of doing things myself. 现在 solo build、用 agents 完成工作时，没有这些 bottlenecks；唯一 bottleneck 就是我自己，我自己的 work 限制了一切。

嘉宾 So I don't have these bottlenecks uh with like running agents and getting work done uh the way I do now when I build solo. uh when I built solo there is no bottleneck other than myself right everything is like bottlenecked on myself on my own work um so I started running into all these kind of problems how do I validate AI generated code how do I really plan with AI more interactively instead of like looking at the terminal like a long wall of text how do I really juggle through all the 20 30 sessions without going crazy right I started running into these problems and then I I was forced to develop this tooling because there's nothing else that can solve it very well. 因此我开始遇到这些问题：怎样验证 AI-generated code？怎样与 AI 更 interactive 地规划，而不是盯着 terminal 中的一墙文字？怎样协调 20、30 个 sessions 而不发疯？我被迫开发 tooling，因为没有别的东西能很好解决。然后我意外地… …成了 open-source maintainer。现在要处理 open source issues 和 PRs。

So then I uh kind of accidentally became an open- source uh maintainer. Uh now I have to deal with the open source uh uh issues and PRs which is good and it keeps the tooling become better and uh now I have to balance my time. Uh how much do I spend open source maintenance versus uh my original plan of building the uh the apps I want to build. I see. Okay, broad question. What types of software you expect to be changing and what do you think will stick around? 好处是 tooling 会持续改善，但也必须平衡时间：多少用于 open source maintenance，多少用于最初计划的 apps。主持人：一个宽泛问题：你预计哪些 software 会改变，哪些会保留下来？ Kun：我认为有几个大 trends。过去二十年构建的大量 software，尤其 SaaS，主要面向 humans，本质上是 human interface 加后端 database，再用某种可扩展 system 管理。

Yeah. Yeah, that's a really good question. I think there's a few big trends that are happening. One is that a lot of the software uh especially the SAS software that were built in the past two decades are mostly built for humans, right? They are essentially a human interface plugged in with a database in the back end and have a somewhat scalable system to manage that. 我认为 human interface 这一部分会基本消失。我们不想自己在 Salesforce 等 SaaS websites 中一路点击。

The human interface part I think will pretty much like go away. we don't really want to click through all the websites ourselves in like in Salesforce in uh all those SAS uh websites right most of the time we want our agents to work with those services and agents need a different interface to work with that so I think there is going to be a rebuild a very big rebuild of a lot of these services to really make them more headless and uh get rid of the human interface aspect of it or change the human interface is to be something that allows the human to collaborate with AI to get work done. 多数时候，我们希望 agents 与这些 services 合作，而 agents 需要不同 interface。因此大量 services 会经历一次大 rebuild，变得更 headless，移除 human interface，或把它改成 humans 与 AI 协作完成工作的界面。我认为这场大规模重建会发生。能为 agents 提供优秀 interfaces、让它们高效操作 services 的 software 会留下。

Right? So there's a big rebuild that I think will be happening. The software that gives agents really good interfaces so that agents can work very productively with those services those I think will remain. So right now I think if we think of the new kind of like GitHub equivalents I think whoever that will win that race to become the new GitHub has to nail this aspect. 比如思考新一代 GitHub equivalent，最终胜出的产品必须把这一点做好：它应是 agents 可以操作的 headless service。而不是需要 humans 点击的 website。主持人：这具体是什么？更好的 documentation、稳定 API schema，还是别的？

It needs to be a headless service that agents can operate um not like a website that humans have to go click through. So what does that look like? Is it like better documentation, stable API schema? Like what what are the things that actually matter for agents? Yeah. Yeah. So this is something that's also very interesting. Let me bring up this site I have called axi.md. So axi this is something I discovered I started to get into when I started building tools for agents. 对 agents 真正重要的是什么？ Kun：这也非常有意思。我打开一个叫 axi.md 的 site。 AXI 是我开始为 agents 构建 tools 时逐渐研究的东西。我发现很多 tools，包括 MCP servers 和 CLIs，都没有真正针对 agents 优化。

I noticed that a lot of the tools including MCP servers and uh COIs that people developed, they were not really optimized for agents. Um so take uh MCP servers for example. Um I think a lot of companies are still learning how to develop really good interfaces for agents. Uh so as a result, some of the MCP servers are really not very efficient. GitHub, I will probably call them out. 以 MCP servers 为例，许多 companies 仍在学习怎样为 agents 开发优秀 interface，结果有些 MCP server 非常低效。我要直接点名 GitHub。 GitHub MCP server 大概是最无效、最没必要的 MCP server，因为 GitHub 已经有非常好的 CLI，agents 能通过它完成多数工作。

Uh, GitHub MCP server is probably the most inefficient and unnecessary MCP server out there because GitHub has a really good COI and agents can use GitHub COI to get most of the work done. Why do we need a MCP server? And uh what I did was that I actually ran out of benchmarks. So, let me see if I can find the GitHub uh benchmark um that we can uh look at here. Here I ran a benchmark to really evaluate for the same tasks, right? 为什么还需要 MCP server？我实际跑了一组 benchmarks，来看看能否找到 GitHub benchmark。对于相同 tasks，只改变 interface 会产生多大影响？

If we just change the interface, how much does that matter? And here we can see uh if we use the GitHub CLI, this is GitHub uh measuring the GitHub operations. If we use the GitHub CLI, this is the average cost. If we use the MCP server, um there's a few variants without two search with two search with code mode, but it's always more expensive than COI to get the same work done. 这里衡量 GitHub operations：使用 GitHub CLI 时有一个 average cost；使用 MCP server 时，无论是 without tool search、with tool search 还是 code mode，都比 CLI 完成同一任务更昂贵。

So this is like uh GitHub operations. uh it's u um this uh MCP based approaches are slower, take more turns, uh spend more tokens. Uh it's just worse in every way. The the GitHub COI is more optimized and it's better I think because uh they have existed for a long time and the LLM's training data already has a lot of examples for how to use that very efficiently. Um so that's why um it's already pretty good. MCP approaches 更慢，需要更多 turns，也花更多 tokens，各方面都更差。 GitHub CLI 优化得更好，原因之一是它已存在很久，LLM training data 中有大量高效使用示例。因此它本身已经很好。对 GitHub operations，除了一些 server-side integrations，几乎没有理由用 MCP。

So I think for GitHub operations uh there is pretty much like no reason to do MCP except for some servers server integrations. So the thing about CLIs is that most of the COIs out there are designed in a human uh development kind of scenario, right? It's humans running those COIs. Yeah. So there's I I so I thought there's probably like room to improve. 但多数 CLI 是按 human development scenario 设计的，由 humans 运行，因此还有改进空间。

So I developed a bunch of uh COIs for agents and I started to really distill down a few principles that can make COIs very ergonomic for agents. Um and my uh proof point is here. The proof point I have here um besides the GitHub benchmark uh where uh I built an axi uh following the principles I have uh to build agent ergonomic COIs uh it can be even cheaper than the GitHub CLI uh and have higher success rate. That's one example. 我为 agents 开发了一批 CLIs，并提炼出一些让 CLI 对 agents 更 ergonomic 的 principles。除了 GitHub benchmark，我还有一个 proof point：按这些 principles 构建的 AXI… …可以比 GitHub CLI 更便宜，同时 success rate 更高。

嘉宾 I also did one for the ChromeDev tools because uh browser automation is also a very very common use case, right? Uh a lot of people um use agents uh and run their browser to do various kind of things. So I benchmarked very kind various kind of common uh approaches to connect browser to the agent. There's uh chrome dev tools empty mtp server which is very popular. There's agent browser. There's dev browser. There's a few tools like that. 我也为 Chrome DevTools 做了一个，因为 browser automation 是很常见的 use case。许多人让 agents 操作 browser 完成各种事情。我 benchmark 了几种常见 browser-agent connection approaches，包括很流行的 Chrome DevTools MCP server。

Um I built a chrome dev tools axi by following the principles I defined uh and like iterated on over time following those principles and I built a wrapper. This chrome dev tools axi is a wrapper of the chrome dev tools mcp. So there's no functional difference. It does everything the same way. It's like the MCP server under the hood. I just only changed the interface and now it's significantly more uh more efficient. 还有 Agent Browser、Dev Browser 等。我按照定义并持续迭代的 principles 构建 Chrome DevTools AXI。它只是 Chrome DevTools MCP 的 wrapper，功能完全相同，底层仍使用 MCP server，我只改变 interface。

So the average cost reduced by over 20%. Without losing anything and latency also uh improved a lot. So that's like uh the browser use case. The principles I uh really distilled this down to are 10 principles here. Uh so I documented everyone can go to the axi.mmd website to see this. Um it's just like a public set of standards. The first one is like use token efficient output. A lot of the MCP servers and uh COI tools they use JSON to output data. 结果效率显著提升，average cost 降低超过 20%，没有损失任何功能，latency 也改善很多。这是 browser use case。我最终提炼出十条 principles，并公开记录在 axi.md website。

Right? So JSON is very useful when you have something else that needs to parse the data in a structured way. But LLM's agents don't parse the data like that. Right? They only need to understand the content and the semantics of the data. So use the token efficient output format. Uh it can save a lot of tokens straight out of uh out of the gate. And then there's like minimal default schema. 第一条是使用 token-efficient output。很多 MCP servers 和 CLI tools 用 JSON 输出 data。 JSON 很适合交给其他系统做 structured parsing，但 LLM agents 不会这样解析；它们只需要理解内容和 semantics。

Uh some like CI tools when you ask for some data it will produce like literally every single column every single field. So various kind of things like that I documented them as a set of principles that can make an interface really ergonomic for agents. And by following this uh I was able to like uh consistently build rappers and tools that are just more efficient for agents for uh than their original counterparts. And I have a few published over here. 因此使用 token-efficient output format，开局就能节省大量 tokens。另一条是 minimal default schema。有些 CLI tools 请求一点 data，却把每一列、每一个 field 全部输出。我把各类问题整理成一组 principles，让 interface 真正对 agents ergonomic。

Uh there's a catalog here as well. I have GitHub axi, chrome dev tools axi, lavish itself is an axi. I recently published a quotota axi as well that can expose the agent quotota uh data to your agents. This is for first mate to be smart about not using a crewmates that are like not using a subscription that's mostly out of kota. And then there's a bunch of community contributed uh axis already. 遵循它们后，我能够持续构建比原始 counterpart 更高效的 wrappers 和 tools。这里已经发布了几个，也有 catalog：GitHub AXI、Chrome DevTools AXI；Lavish 本身也是 AXI。

Uh there's uh npm sqlite a lot of the common tools that we use already have an axi version. So, uh I definitely recommend uh people who like uh are running into the bottleneck of token efficiency to start looking into like just using more efficient tools. Real quick, if you want to implement Coon's entire setup for yourself, it's going to be available in the second link below the video completely for free. 我最近还发布了 Quota AXI，把 agent quota data 暴露给 agents，让 FirstMate 更聪明，不会选择 subscription quota 几乎耗尽的 crewmate。社区也已经贡献了许多 AXIs，包括 npm、SQLite 等常用 tools。

This includes his first made setup herder Westerm and everything else uses to actually build with AI agents. Again, grab it completely for free. Second link below the video. Man, this is amazing. like I'm going to run all my softwares through this to make sure they're like agent native. Yeah, I I could like talk about this for hours. Let me ask on the like product level. 非常建议遇到 token-efficiency bottleneck 的人开始使用更高效的 tools。主持人：如果想完整复刻 Kun 的 setup，视频 description 第二条链接完全免费，包含 FirstMate、Herder、WezTerm 及其他全部工具。

Basically, you know, I think me and you probably agree that like soon enough 99.9% of software and tools will be used by agents. And again, we will talk to our main agent, you know, whatever name we have for that. Yeah. Why are so many people still missing this? Because so many people still are like focused on like, oh, I'm going to build this like human native web UI and they don't think about the back end at all. they don't think about the CLI at all. 再次说明，可免费获取。主持人：太棒了。我会把所有 software 都通过它检查一遍，确保足够 agent-native。这个话题我能聊几个小时。换到 product level：我们大概都同意，很快 99.9% 的 software 和 tools 都会由 agents 使用，而我们只与自己的 main agent 对话。

What what was the moment you realized like agents are going to run all the software? We're just going to talk to like our main agent. That's a good question. I think um this is a learning curve. We have seen this repeated every time when there is a new technological transformation. Uh so all the industrial revolutions, right? 为什么仍有这么多人没意识到？很多人还专注构建 human-native web UI，完全不考虑 backend 或 CLI。

uh when like uh steam engines were introduced, when electricity was introduced, when internet was happening, uh it always started with a small set of early adopters who are just really uh who just really like to tinker around,right? Uh who who likes to play with the technology even if the technology is not paying off. So I think we're at that phase right now where a lot of the tinkering in some areas is indeed not going to pay off. 你在哪个时刻意识到 agents 将运行全部 software，而我们只会与 main agent 交流？ Kun：这是很好的问题。我认为这里存在 learning curve。每次发生 technological transformation 都会重复同一模式。

Um so there are still a lot of legacy projects where AI can only help you so much. It can still help but it's not going to be as big of a boost as uh what you get from a green field projects right. 所有 industrial revolutions，无论 steam engine、electricity 还是 internet，起初都由一小群 early adopters 推动，他们就是热衷 tinkering。

Um so people who are in those uh domains when they try AI they will see oh it's like not that helpful. why do I spend the effort to really um work with it and those are reasonable arguments but that's the reason we we see um the early adopters are usually tinkerers who are just playing with the technology fake for the sake of it but those are the people who will figure out what is actually going to work those are the people who will figure out the tooling the gaps and uh ways of working that can really benefit from this new technology and those people will start to share those tools and products with others And this takes time. 即使技术还没有带来回报，他们也愿意玩。我们现在就处于这个阶段，某些领域的许多 tinkering 确实不会获得回报。很多 legacy projects 中，AI 的帮助仍有限；它有用，但提升不如 greenfield projects 那么大。身处这些 domain 的人尝试 AI 后会觉得“没那么有用，为什么还要投入精力”，这是合理观点。但正因如此，early adopters 往往是为了技术本身而玩的 tinkerers；他们会弄清什么真正可行。

So it takes time for the early adopters to really prove something is good. Uh it takes time for them to uh boil down those uh those experiences into uh really good tools that others can also use. 他们会找出 tooling、gaps 与能真正从新技术获益的工作方式，再把 tools 和 products 分享给别人。

It takes time for others to really get convinced of the new tools and start to spend time with it. uh so all this is going to take time and I I think it's reasonable to expect that it's not going to be overnight that everyone will realize the same thing that uh AI is uh is here and this is the way uh everything will change um I think everyone is uh in a different circumstance and their constraints what they do every day uh the projects they're working with etc etc will shape what's the right timing for them to get through that uh inflection point I think this is Amazing point to end the solon. 这个过程需要时间：early adopters 要先证明某种方法很好，再把 experiences 提炼成其他人也能使用的优秀 tools。其他人也需要时间被新 tools 说服，并愿意投入。所以不会在一夜之间让所有人意识到 AI 已经到来、一切将因此改变。每个人的 circumstances、constraints、日常工作和手中 projects 都不同，会决定他们何时跨过自己的 inflection point。

I couldn't agree more. Thank you very much for spending your time. What are the main things people should go and check out? Maybe check out my GitHub uh where I listed a lot of the tooling that I uh I built and shared. I pretty much open source everything I have. Uh so everything uh I I do in my workflow that helps me get a lot done are already shared on my GitHub uh repo. 主持人：这是非常适合收尾的观点，我完全同意。感谢你花时间分享。大家最应该去看什么？ Kun：可以看看我的 GitHub，那里列出了我构建和分享的大量 tooling。我几乎把所有东西都 open source 了，workflow 中帮助我高效完成工作的内容都已放在 GitHub repos。

And I also have a uh YouTube channel where uh I uh made some videos walking through exactly how I code, how I use the tools that I built to uh build new things efficiently. Uh so those are really uh good resources as well. Those are probably good starting points. 我还有一个 YouTube channel，视频会完整展示自己怎样 coding，以及如何用自建 tools 高效构建新东西。

Um, and if you run into any problems using the tools or have questions about how to use agents more efficiently, I have a Discord channel as well, uh, Discord server, uh, where, um, we have a pretty good community of really people who are really helpful, uh, at, uh, helping each other, uh, and I often go there to discuss with everyone as well. Awesome. I'm going to link all of that below. Cool. Cool. Thanks, David, for having me here. Yeah, likewise. 这些都是不错的起点。如果使用 tools 遇到问题，或想了解怎样更高效地使用 agents，我还有… …一个 Discord server，里面有很好的 community，大家积极互相帮助，我也常去与所有人讨论。主持人：太好了，相关链接都会放在下方。 Kun：感谢 David 邀请。主持人：我也很感谢。