我们为什么从 Claude Code 切换到 Codex

Codex is one of those things where three months ago, six months ago, it was trash. Codex 这东西，三个月前、六个月前都还很烂。 If anyone from OpenAI is on the call and listening to that, I stand by that 100%. 如果现在有 OpenAI 的人在听，我一百分之百维持这个说法。 If you have a great generalpurpose coding agent on your computer, it's actually really great for any kind of knowledge work. 如果你的电脑上有一个优秀的通用编程智能体，它其实对任何知识工作都非常有用。 If it can write software on its own, it can do any kind of knowledge work on its own. 它能自己写软件，就能自己做任何知识工作。 When I sign on during the day, Codeex is the first thing I open. 每天开电脑，Codex 是我打开的第一件事。 It is pulling in whatever I need from Gmail, Slack, [music] Notion, Stripe, all of our data sources. 它会帮我从 Gmail、Slack、[音乐] Notion、Stripe，以及我们所有的数据源里拉取我需要的信息。 It's where I spend like 80% of my time working overwhelmingly because the app itself is just so good. 我 80% 的工作时间都花在这里，因为这个 app 本身实在太好用了。 There's a new operating system for how and where you're going to get your work done and it's this kind of agent management interface. 知识工作的方式和场所正在迎来全新的操作系统，那就是这种智能体管理界面。 Hello everybody. 大家好。 Welcome to Codeex Camp. 欢迎来到 Codex Camp。 Codeex for knowledge work. Codex 与知识工作。 Psyched to have you. 很高兴有你们在。 Psyched to have you on this auspicious GPT 5.5 day after release day. 很高兴能在 GPT-5.5 发布次日这个吉利的日子里跟你们见面。 Hope you're doing well. 希望大家一切都好。 Um I'm here with our head of growth, Austin. 我旁边是我们的增长负责人 Austin。 Austin, say hello. Austin，跟大家打个招呼。 Hello. 大家好。 We're psyched to have you. 很高兴有你在。 We are psyched to do this. 我们都很期待这次对话。 Um Codex is one of those things where, you know, three months ago, six months ago, it was trash. Codex 就是那种，你知道的，三个月前、六个月前，还是一坨垃圾的东西。 Um, and if anyone from OpenAI is on the call and listening to that, I stand by that 100%. 如果现在有 OpenAI 的人在听，我一百分之百维持这个说法。 Um, and it was really built for um, senior engineers uh, doing pair programming. 它最初是专门为高级工程师做 pair programming 打造的。 So, it was um, it would argue with you, it would make you feel stupid. 所以，它会跟你争论，让你觉得自己很蠢。 It was just it was like a little autistic like it it didn't have any emotional intelligence. 它就是——就像有点神经质，完全没有情商。 And I think OpenAI had this interesting strategy or this interesting theory starting with GBT5 that your vibe coding was going to happen in chatbt and that was where all that stuff was going to live and then senior engineers are going to use codecs to like do all their programming work but we're going to hobble the model so it doesn't do anything bad. 我觉得 OpenAI 从 GPT-5 开始有一个很有意思的策略——他们认为 vibe coding 会发生在 ChatGPT 里，所有那些事情都在那里完成，然后高级工程师再用 Codex 做全部的编程工作，但他们要给模型套上枷锁，不让它做任何出格的事。 It's in a sandbox all that kind of stuff. 放在沙箱里，诸如此类。 And I think basically what happened is Anthropic figured out that having a model that's pretty usable and fast and smart and also emotionally intelligence intelligent on your computer that can access your computer um is a really really great experience for programmers. 我觉得，Anthropic 基本上发现了一件事：在你的电脑上有一个能用、够快、够聪明、还有情商、能访问你电脑的模型，对程序员来说是一种真正了不起的体验。 And it means you could throw away a lot of the old uh stuff that you used to have in a in a programming environment where you it was built for typing code. 这意味着你可以扔掉编程环境里很多老旧的东西——那些过去专门为敲代码而生的玩意儿。 You could just type commands into your terminal and then it would start working. 你只需要在终端输入命令，它就开始干活了。 And then I think what Anthrobic figured out is if you have a great general purpose, if you have a great coding agent on your computer, it's actually really great for any kind of knowledge work. 我觉得 Anthropic 接着发现，如果你电脑上有一个优秀的通用、优秀的编程智能体，它其实对任何知识工作都非常有用。 If it can write software on its own, it can do any kind of knowledge work on its own. 它能自己写软件，就能自己做任何知识工作。 And we started to move from this world where programmers had been delegating um had been delegating their tasks starting to delegate their tasks inside of cloud code to now any kind of knowledge work is being delegated inside of cloud code and cloud co-work and all that kind of stuff. 我们开始从那个时代走出来——程序员曾经只是在 Claude Code 里把任务交给它，现在任何知识工作都可以在 Claude Code 和 Claude Co-work 之类的工具里被委托出去。 And I think openai they had this original split. 我觉得 OpenAI 最初的想法是把两者分开的。 It's like oh you're going to do all your vibe coding in chatbt and I think they saw what was starting to happen with cloud code and over the last maybe three months or so they have done this hard pivot on co codeex where it has gone from a senior engineer only tool that is really for pair programming um to I think like it's it is my daily driver for this kind of work um I uh I I use codeex for everything from deep engineering stuff to writing to recruiting. 他们的思路是：你在 ChatGPT 里做 vibe coding，然后他们看到了 Claude Code 那边开始发生的事情，在过去大概三个月里，他们对 Codex 做了一次硬转型——从一个只适合高级工程师做 pair programming 的工具，变成了我现在这类工作的日常驱动；我用 Codex 做从深度工程到写作再到招聘的所有事情。 I do a lot actually do a fair amount of recruiting. 我确实做挺多招聘的，还真不少。 It's really good for that and I'll give you some use cases um later. 这方面它很厉害，之后我会给大家举几个例子。 But they [snorts] sort of figured out that um having this general purpose agent on your computer with the ability to write code, the ability to access your file system, the ability to have a browser um and wrapping it in a desktop app is like the ideal uh ideal next step for knowledge work. 但他们 [嗤笑] 基本上发现，让这个通用智能体待在你的电脑上，具备写代码的能力、访问文件系统的能力、使用浏览器的能力，再把它包装成一个桌面 app，是知识工作下一步最理想的形态。 And I think that they built the best current version of that. 我觉得他们打造了目前这个形态里最好的版本。 Um, and it what it is starting to snap into into focus now is that there's a new operating system for how and where surface for how and where you're going to get your work done. 现在开始浮现的一个清晰轮廓是：知识工作将在一个全新的平台上完成，而那个平台就是这种智能体管理界面。 And it's this kind of agent management interface. 也就是这种智能体管理界面。 And that's whether or not you're using cloud code or cloud co-work in the desktop app or codeex in the desktop app. 无论你用的是 Claude Code、Claude Co-work 桌面 app，还是 Codex 桌面 app，都是这个方向。 It's becoming this race between the model companies where every each model company has their own surface like this for agent management, a desktop app for agent management that's at its core a programming agent that's used for knowledge work. 这正在演变成模型公司之间的竞争，每家模型公司都有自己的这类桌面 app，核心是一个用于知识工作的编程智能体。 Um, Anthropic has cloud code and cloud co-work. Anthropic 有 Claude Code 和 Claude Co-work。 OpenAI has codeex. OpenAI 有 Codex。 XAI recently um essentially bought cursor um and uh and Google is the only one that I mean they have anti-gravity but I don't think no one is seriously using it for that yet but I I imagine Google will do this too and that's the race that is the race that's happening and so I think for us who gets who get all the benefits of uh being able to use these tools uh it's really important to uh be be bouncing around between these. xAI 最近基本上把 Cursor 买了下来，Google 是唯一一个——我是说他们有 Anti-gravity，但我觉得还没人认真用它做这个，不过我想 Google 迟早也会入场，这就是正在发生的竞争。所以对我们这些能用上这些工具的人来说，在它们之间来回切换非常重要。 So like using for example using codecs so that you can feel what it's like to work in an agent first world because once you add once you add an agent that is like the your primary way of accessing and using software and the internet and all that kind of stuff, it opens up all this interesting stuff that wasn't possible before because you can send your agent out to go talk to other pieces of software and come back and um you know we can get into into more of the details there but I want to get into like the more of the concrete use cases but that's the world that we're starting to live in. 比如用 Codex，去感受在一个以智能体为核心的世界里工作是什么感觉，因为一旦你把一个智能体作为访问和使用软件、互联网以及一切的主要方式，就会打开很多以前不可能的东西——你可以把智能体派出去跟其他软件对话再回来汇报，我们可以深入聊聊细节，但我想先聊具体用例，总之我们正在进入这个世界。 You're doing work on your computer through codecs or co-work. 你在通过 Codex 或 Co-work 在电脑上工作。 And um and your agent is your interface to a lot of the work that you're doing and a lot of the a lot of the software that you use and a lot of the stuff that you do every day. 你的智能体成为你完成很多工作、使用很多软件、处理很多日常事务的接口。 And uh that's actually really fun. 这其实真的挺好玩的。 It's really cool. 真的很酷。 There's a lot of good stuff here. 有很多好东西。 And so I wanted to uh I wanted to bring Austin in to to help do this because Austin is our head of growth and I think he had his real agent pill moment. 所以我想把 Austin 请进来，因为 Austin 是我们的增长负责人，我觉得他有过真正意义上的「智能体觉醒」时刻。 You tell me Austin, but probably like three or four months ago and the agent pill moment was really cla on a on a Monday morning being like, Austin，你告诉我，大概是三四个月前，那个觉醒时刻——真正的觉醒就是某个周一早上—— "Oh yeah, I just was on my computer all weekend. 「对，我整个周末都在电脑前。 Like I I was like 12 hours a day. 就是，我一天坐了大概 12 个小时。 didn't go out or anything because I was using cloud code and um you started to use it for all those all the kind of knowledge work tasks that a that a growth marketer would and over the last couple weeks as we've been using 55 and I've been telling you for a little bit you should try codeex it seems like you've you've actually just shifted everything over to codeex and 55 and so I think you're a great person to talk about you know sort of what you're seeing and how and how that is uh how that is um how how this has changed how these agent management interfaces have changed your workflow and then why you like codecs and then I would love to get into some demos of your actual codec workflow so that we can sort of see things uh from your perspective. 哪儿也没去，因为我在用 Claude Code。然后你开始把它用在增长营销人员会做的各种知识工作上，过去几周我们一直在用 GPT-5.5，我也一直在跟你说要试试 Codex，看起来你已经把所有东西都切到 Codex 加 GPT-5.5 了。所以我觉得你是个很好的人来聊聊这些：你看到了什么、这有多大影响、这些智能体管理界面如何改变了你的工作方式、为什么你喜欢 Codex，然后我很想看看你实际的 Codex 工作流 demo，这样我们就能从你的视角直观地看到一切。 Yeah, that sounds great. 好，听起来不错。 So I um yes, my like agent pill moment was spending a week going deep into cloud code in the CLI uh probably in like December into January, hooking it up to everything I do for work and for my personal life and finding that I I use Warp as my like CLI interface. 所以，是的，我的「智能体觉醒」时刻就是——大概在十二月到一月——花了一个礼拜深度使用 Claude Code CLI，把它接入了我所有的工作和个人生活，我用的是 Warp 作为 CLI 界面。 um and finding that the things it could automate, the things it could handle for me, and then the way it could work as a thought partner to make my work better. 然后发现它能自动化的事情、能替我处理的事情，以及它作为思考伙伴让我工作更好的方式。 It was like this is the only way I want to do the kind of knowledge work that requires um strategic thinking and uh data analysis and shipping marketing copy, like a bunch of stuff that can get you spread out across a bunch of apps and tools during the day. 就感觉——这才是我想要的知识工作方式，那些需要战略思考、数据分析、撰写营销文案的工作，就是那种让你在一天里横跨一堆 app 和工具的事情。 And in maybe in February, you you kept nudging me to be like,"You really should try Codeex. 大概二月的时候，你一直在推我：「你真的应该试试 Codex，有些东西你会喜欢的。」 There were things you liked about it." 有些你会喜欢的东西。」 And if someone says that at every if anyone on the team says that, like, I'll go try it. 如果团队里有人这么说，我肯定去试。 And I like to push myself and play around with more engineeringy tasks, especially to see what these models are capable of. 我喜欢推着自己去尝试更有工程感的任务，尤其是为了看看这些模型能做到什么。 And so I tried to build a personal vioded app in Codeex because that was one of the things that you said that it was really good for. 所以我尝试在 Codex 里做一个个人视频 app，因为那是你说它特别擅长的事情之一。 And my immediate response was like I think it is better at building the app, but I can't tell because it's nothing has ever made me feel more stupid than codeex like two months ago. 我的第一反应是：我觉得它在构建 app 这件事上更厉害，但我说不准，因为从没有什么东西像两个月前的 Codex 那样让我觉得自己蠢透了。 Like I always I use compounded our compound engineering plugin that Kieran Classen made for basically everything including knowledge work, but especially if I'm trying to build an app or ship a PR to the to the site. 我一直用 Kieran Classen 做的 compound engineering 插件，几乎做所有事情，包括知识工作，尤其是在我想构建 app 或给网站提 PR 的时候。 So I made a plan. 所以我做了一个计划。 in the plan it comes up with like three questions and uh for like which direction we should go and uh I had no idea what the hell it was talking about. 计划出来之后，它提了大概三个问题，问应该往哪个方向走——我完全不知道它在说什么。 It was like do you do one of any of these three and every question? 就像「这三个方向你选哪个？」，每个问题都是这样。 And I was like,"Please explain this to me um in more detail." 我说：「能解释得更详细一点吗？」 And his response was basically like,"Why?" 它基本上的回应就是：「为什么？」 Like,"Why don't you just do what I'm recommending?" 就是「你为什么不直接照我推荐的做？」 And I found a way to I I basically stayed in codeex for all engineering stuff because I I did like the results even if I didn't love working in it. 我找到了一种方法——我基本上留在 Codex 里做所有工程相关的事，因为我确实喜欢结果，即使不太爱在里面工作。 But I would say 80% of what I was reaching for was was cloud code in the CLI. 但我 80% 时间伸手去拿的还是 CLI 里的 Claude Code。 And when we got our hands on the new GPT model a month ago, the the the the first thing I felt was at the very least there's parody between the latest Opus model and the latest GPT model for the kind of knowledge work I do. 后来我们提前拿到新 GPT 模型大概一个月，我的第一感受是——至少在我做的这类知识工作里，最新的 Opus 模型和最新 GPT 模型之间已经平起平坐了。 There's a few things that Opus does better. Opus 有几件事做得更好。 There's a few things that that Codeex does better. Codex 有几件事做得更好。 That feels a little more specific to me even like I I outside of design, which I still really trust Opus for. 在我眼里感觉更具体一点，比如设计——我还是非常信任 Opus 的。 um it feels a little more like, oh, there's some stuff I like better than this. 感觉更像是：「哦，这个有些东西我更喜欢。」 But the real differentiator to me is that to me there's no comparison for how fast and powerful the codeex desktop app is as just like an app compared to the claw desktop app. 但真正的差异化因素对我来说是：Codex 桌面 app 作为一个纯粹的 app，跟 Claude 桌面 app 相比，速度和功能完全不在一个量级。 Like I have never been able to get uh co-work to work for me. 我从来没能让 Co-work 真正适合我。 And I think it's because I've been kind of ruined by the codeex app. 我想是因为我已经被 Codex app 惯坏了。 It's so fast. 它太快了。 The sub agents are so good. 子智能体太好用了。 the way in which it suggests and then um ships automations for me is just like it I can't imagine not using it. 它为我推荐和交付自动化的方式——我真的无法想象没有它的工作。 I wouldn't be surprised if any week the cloud desktop app is like just as good, right? 我不会惊讶哪天 Claude 桌面 app 一下子追上来，对吧？ like um they could ship versions where it's faster and better, but I I'm now at the point where when I sign on during the day, I codeex is the first thing I open. 他们可以发布更快更好的版本，但我现在已经到了每天开电脑，Codex 是第一个打开的东西这个状态。 It is pulling in whatever I need from Gmail, Slack, Notion, Stripe, all of our data sources. 它会帮我从 Gmail、Slack、Notion、Stripe 以及我们所有的数据源里拉取我需要的信息。 This morning I was like,"Oh yeah, we need to do a run of show for this camp." 今早我就想：「对，我们需要为这次 Camp 做一个 run of show。」 I messaged Codeex. 我发消息给 Codex。 I'm like,"Make the run of show." 我说：「做个 run of show。」 It knows exactly where to look because we've already had conversations about what we're going to talk about today. 它准确知道去哪里找，因为我们之前已经聊过今天要讲什么了。 It pushed it to notion. 它把内容推送到了 Notion。 It sent it to Slack. 发到了 Slack。 It was perfect. 完美无缺。 It was like,"Oh, yeah. 就是那种感觉：「对， This is exactly what we should do." 这就是我们该做的。」 And um yeah, it's where I spend like 80% of my time working overwhelmingly because the app itself is just so good. 我 80% 的工作时间都花在这里，因为这个 app 本身实在太好用了。 And then the model has now gotten good enough to be the daily driver for me. 然后这个模型现在也足够好，已经成为我的日常驱动了。