How I manage 15 AI agents 24/7 as a solo founder | Ryan Carson
频道: How I AI
视频: https://www.youtube.com/watch?v=zPfxlcVpFgs
原文语言: en
统计: 共 200 轮 · Ryan 97 · Claire 101
[0:00] Ryan
If you're out there listening and [music] you are doing engineering, you're doing your work locally, you really need to open your eyes. I think the future is pretty much 100% cloud agents. I went all in. So much so, I spent $5,000 on Devon and then I went up to $20,000 one month. You know, I think one of my big messages for today is that all of us have to uplevel our ability to manage agents. Like that is our job. How do you effectively manage 10 20 100 500 employees? This is the whole reason that organizations were built as pyramids like humans can't scale and have a thousand directs.
如果你正在听这期节目，而你是做工程的、活儿还都在本地干，那你真得睁开眼看看了。我觉得未来基本上会是 100% 的云端 agent。我是全押进去了。押到什么程度呢——我在 Devin 上花了 5000 美金，后来有一个月直接花到 2 万。我觉得我今天最想传达的一点是：我们所有人都得把「管理 agent」这个本事提上一个档次。这就是我们现在的工作。你要怎么高效地管 10 个、20 个、100 个、500 个员工？组织之所以被设计成金字塔，根源就在这儿——人是没法一个人带一千个直属下属的。
[0:37] Claire
What is in your stack right now and how have you structured stuff to make you successful and build things effectively?
你现在的技术栈里都装了什么？你是怎么把这一套搭起来、让自己能高效出活、把东西真做出来的？
[0:43] Ryan
So what I'm doing is figure out how to manage you know throngs of agents and in Devon what what you're seeing is uh is folders and you'll notice P 0 is the big stuff. the these are the the main things I need to move forward as the business. New features, [music] P1, P2. So at priority zero, I've got bugs. Let them work. Usually give them a goal. And it's just such a crazy different world we live in now. [music] Welcome back to How I AI. I have a super special guest, Ryan Carson, one of our earliest How I AI episodes who is here to show us what has changed [music] in the last 12 months in the age of coding agents. We're going to wax poetic about our favorite little AI intern, Devon. Talk about all the ways you're underutilizing your coding agents, and give maybe the most anti- AI episode of How I AI, where we talk about where AI can't solve your problems, how you need to get out in front and actually talk to people, and how pen and paper still have a place in his stack. Let's get to it. This episode is brought to you by WorkOS. [music] AI has already changed how we work. Tools are helping teams write better code, analyze customer data, and even handle support tickets automatically.
我现在在琢磨的事，就是怎么管住一大群 agent。你在 Devin 里看到的这些就是文件夹，你会注意到 P0 装的是最要紧的事——是推动业务往前走的主线，比如新功能。然后是 P1、P2。优先级 0 那一档里我放的是 bug。让它们自己干就行了，我一般就给个目标。我们现在待的这个世界真的太不一样了。欢迎回到 How I AI。今天请到一位超级重磅的嘉宾，Ryan Carson——他是 How I AI 最早几期的嘉宾之一，这次回来给我们看看，在 coding agent 的时代里，过去 12 个月都变了什么。我们会好好聊聊我俩最爱的那个 AI 小实习生 Devin，聊聊你有哪些地方还没把 coding agent 用到位，然后大概会做出一期最「反 AI」的 How I AI——我们会聊 AI 解决不了你哪些问题、你必须自己走出去跟真人聊，以及为什么纸和笔到今天仍在他的工具栈里占着一席之地。开始吧。本期节目由 WorkOS 赞助。AI 已经改变了我们的工作方式，各种工具正在帮团队写出更好的代码、分析客户数据，甚至自动处理支持工单。
[1:59] Claire
But there's a catch. These tools only work well when they have deep access to company systems. Your co-pilot [music] needs to see your entire codebase. Your chatbot needs to search across internal [music] docs. And for enterprise buyers, that raises serious security concerns. [music] That's why these apps face intense IT scrutiny from day one. To pass, they need secure authentication, [music] access controls, audit logs, the whole suite of enterprise features. Building all that from scratch, it's a massive lift. That's where work OS comes in. Work OS gives [music] you drop-in APIs for enterprise features so your app can become enterprise ready and scale [music] up market faster. Think of it like Stripe for enterprise features. OpenAI, Perplexity, [music] and Cursor are already using WorkOS to move faster and meet enterprise demands. Join them and hundreds of other industry leaders at workos.com. Start building today. Ryan Carson, number one episode of How I AI from our early early days, is back again. I am so happy to have you here, my friend. Oh, this is it's so fun to reconnect. Um, and the reason why the show is good is because of you. And, uh, just everybody listening, I'm not just kissing ass here. Uh, everybody that I meet, I say there's one show. If you're stressed about keeping up on AI, just listen to Claire's show.
但有个前提。这些工具只有在能深度接入公司系统时才真好用。你的 co-pilot 得看得到你的整个代码库，你的 chatbot 得能检索所有内部文档。而对企业采购方来说，这就带来了严重的安全顾虑。所以这类应用从第一天起就要接受 IT 部门的严格审查。要过这一关，它们需要安全认证、访问控制、审计日志——一整套企业级功能。从零把这些造出来，工作量大得吓人。WorkOS 干的就是这件事。WorkOS 给你一套开箱即用的企业功能 API，让你的产品直接具备企业级能力，更快打进大客户市场。你可以把它理解成「企业功能界的 Stripe」。OpenAI、Perplexity 和 Cursor 都已经在用 WorkOS 提速、满足企业客户的要求。上 workos.com，加入他们和其他几百家行业头部公司。今天就开始搭。Ryan Carson，How I AI 最早期那一期的主角、播放量第一名，又回来了。你能来我真的太开心了，我的朋友。哦，能再聊上真的太有意思了。而且这个节目之所以好，就是因为有你。听众朋友们，我可不是在拍马屁。我遇到的每个人我都会说：只有一个节目值得听——如果你为跟不上 AI 而焦虑，听 Claire 的节目就够了。
[3:25] Claire
Oh, thank you. Well, I'm stressed keeping up with AI. So, I um will take all that stress on behalf of the audience [snorts] and pass on what I know. What I also am so excited about this conversation is, you know, we did we did our first episode a year ago or even a little more than a year ago and like it's so different now.
哦，谢谢。说实话，我自己就在为跟上 AI 而焦虑，所以我干脆替听众把这份焦虑扛下来，再把我知道的都倒出来。这次聊天让我特别兴奋的另一个点是，我们第一期是一年前、甚至一年多以前录的，而现在已经完全不一样了。
[3:43] Ryan
Oh my god,
我的天，
[3:44] Claire
we were such babies.
那会儿我们简直还是小孩子。
[3:45] Ryan
I know. It's it's it's kind of funny like
就是。其实挺好玩的，
[3:48] Ryan
what we chatted about now and it'll be funny what we chat about now in a year, right?
我们当时聊的那些东西——等一年以后再回头看我们今天聊的，肯定也一样好笑，对吧？
[3:52] Claire
Yes. Totally. And for those who have not gone in the archives, uh we did a really really popular episode about how to manage your agents with like it was like pre-sklls even. It was like three-step mark like you invented skills from first principles. It was like three-step markdown file. We were in an IDE which I haven't opened an IDE in like months. Um,
对，完全是。给还没去翻过往期的朋友说一下，我们当时做过一期特别火的，讲怎么管理你的 agent——那时候连 skills 都还没有。当时基本上是三步的……你等于是从第一性原理把 skills 给发明出来了。就是一个三步的 markdown 文件。我们那会儿还在 IDE 里，而我现在已经好几个月没打开过 IDE 了。
[4:17] Claire
we're in an IDE and and we were really you were very uh prescriptive and directive of your agents. You were very handy very handholdy.
我们那会儿是在 IDE 里，而且你对 agent 特别有规定动作、特别爱下指令，手把手管得很细。
[4:27] Claire
And now, you know, I don't know about you, I'll speak for myself. What the agents are doing is none of my business.
而现在呢，你怎么样我不知道，我就说我自己：agent 在干什么，不关我的事。
[4:34] Ryan
No, let them let them work. I usually give them a goal and yeah, it's just such a crazy different world we live in now. It's so exciting. Well, catch us up um because there's also been a lot of changes on your side in the last year. So, what are you up to and how has it, you know, impacted how you think about and work with AI these days?
对，让它们干，让它们自己干。我一般就给个目标，然后……我们现在这个世界真的太不一样了，特别让人兴奋。那你先给我们更新一下近况吧，因为过去这一年你那边也变了很多。你现在在忙什么？这些变化又怎么影响了你现在看 AI、用 AI 的方式？
[4:53] Ryan
Yeah. So, um I basically had this idea to make divorce better. So, I have uh three amazing sisters. We're like best friends. I have a sibling tattoo that's hard to show. It's right there.
嗯。事情是这样，我当时冒出一个念头：能不能把「离婚」这件事做得好一点。我有三个特别棒的姐妹，我们跟最好的朋友一样。我身上还有个兄弟姐妹的纹身，不太好露给你看，就在这儿。
[5:04] Ryan
You know, we talk every week and then uh two of them started to go through pretty gnarly divorce. Um and my parents got divorced, too. So you know I was like divorce h like no one seems to be talking about divorce and AI what why like and so I just started working on that um and was of course empowered by Thor's hammer you know which is the large language models and and I started to build it um and this is when we chatted last you know I was in cursor you know like you said really handh holdy on the agent and since then thankfully it's just blown up um I basically figured out product market fit and everyone who's listening There's probably a ton of startup founders out there. You know, you're trying to find product market fit. You're pivoting fast. You're And I did what all founders do, right? You know, launched a a version of Untangle, um, which is the product, thought it was for consumers, went out, nobody, Turns out, nobody wants to use AI for divorce. They they want lawyers, people. Just I'm just going to say it. Um, and then I basically thought, okay, not PMF. I I'm old enough to know like and not hang on to things, just pivot. And uh I was like, "Okay, I'm going to get a meeting with a lawyer and just see what they think of this thing." And she I met with this amazing lawyer named Renee Bower and I was like, "Do you love this? Do you hate it? Is this like an enemy? Is it a friend?" And she said, "I love it.
我们每周都聊，后来其中两个开始经历相当难熬的离婚。而且我爸妈也离过婚。所以我就想，离婚这事儿……好像没人在把离婚和 AI 放到一起讨论，为什么呢？于是我就开始做这个东西，当然，手里还握着「雷神之锤」——也就是大语言模型。我就动手做起来了。我们上次聊的时候，我还在 Cursor 里，就像你说的，对 agent 手把手管得死死的。从那以后，谢天谢地，这事儿一下就起来了。我基本上算是找到了 product market fit。听众里肯定有一大堆创业者，你在找 PMF，你在快速转型，你在……而我做的就是所有创始人都会做的事，对吧？我发了一个版本的 Untangle——产品就叫这个名字——本来以为是给普通消费者用的，推出去以后，没人用。事实证明，没人想用 AI 来处理离婚，大家要的是律师。我就直说了。然后我就想，好吧，这不是 PMF。我年纪够大了，知道别死抱着不放，直接转就完了。我就想：行，我去约个律师见一面，看看她怎么看这玩意儿。我见到了一位特别厉害的律师，叫 Renee Bower，我问她：你喜欢这个吗？还是讨厌它？它对你来说是敌人还是朋友？她说：我很喜欢。
[6:26] Ryan
You know, I I there's a parallegal shortage. There's this discovery thing we do in divorce is a nightmare. I I I love it and I want to pay for it." And I was like, "Oh, you how much would you pay?" And she said, "A great number." and now we're we're B2B and we sell to family law firms. Um, and it's just exploding. So, I've been trying to hang on to this solo founder. I'm going to do it all myself thing and I'm completely overwhelmed. [laughter]
她说，现在 paralegal（律师助理）严重短缺，而且离婚案里那个 discovery（证据开示）流程简直是噩梦。我太喜欢这个东西了，我愿意为它付钱。我就说：哦？你愿意付多少？她报了一个特别漂亮的数字。所以我们现在做的是 B2B，卖给家事律所。而且业务正在爆发式增长。所以我一直死撑着「独立创始人、什么都自己来」这个路线，然后现在彻底忙崩了。
[6:52] Claire
You and me both. And for folks that can't wait for the punchline, the punch line is Ryan Carson and Clairvo, famous solo founder. I can do it all myself. I need no friends other than AI. [laughter] did our pre-show conversation talking about our regretful and imminent hiring of a bunch of hum
我也一样。给等不及听结论的朋友先剧透：结论就是，Ryan Carson 和 Claire Vo，两位著名的独立创始人、「我一个人全都能搞定，除了 AI 我不需要别的伙伴」的那种人——我们在录制前的闲聊里，正在聊我们那让人后悔、却又迫在眉睫的招人计划，要招一批人
[7:13] Claire
of humans. I'm gonna have to hire humans. And if any of those humans I'm talking to are listening, I love you and I'm excited about working with you now. But but the AI Yeah, it's
一批人类。我得招人类了。如果我正在接触的那些候选人在听这期节目：我爱你们，我很期待跟你们一起工作。不过说回 AI，是这样
[7:23] Ryan
you get spoiled hiring AIs. I mean, you know, you give them 45 pieces of feedback an hour. Um, you know, you're you anyway. So, yeah, going to hire some humans and grow the Untangle team. Very excited. But you haven't yet.
招惯了 AI 你是会被惯坏的。我是说，你可以一小时给它们提 45 条反馈。反正，你就……总之，是的，我要招一些人类，把 Untangle 的团队搭起来。非常期待。不过你还没招呢。
[7:37] Claire
And so I mean the the the good the good end part of the story is you were able to get to rapid iteration, lots of prototypes, product market fit with you and our mutual intern, Mr. Devon.
所以我想说，这故事的好结局是：你靠你自己，加上我们共同的那位实习生 Devin 先生，做到了快速迭代、做了一堆原型，最后找到了 product market fit。
[7:53] Ryan
Mr. Devon.
Devin 先生。
[7:54] Claire
Mr. Devon. And so, you know, when when you were starting out again, you were like really coding, doing requirements, documents, all of that with AI, but you were very human in the loop. And as I've watched you over the past year, I've seen you slowly turn into AI bro who runs agents overnight. Um, so tell us about like what is in your stack right now? What is like very different from what we talked about a year ago? And how have you structured stuff to to make make you successful and build things effectively?
Devin 先生。那我们说回来，你最开始的时候，是真的在写代码、写需求文档，全都用 AI 做，但人在环里管得非常紧。而这一年我一路看着你，眼看你慢慢变成了一个「通宵放 agent 跑」的 AI bro。所以跟我们讲讲，你现在的技术栈里都装了什么？跟一年前我们聊的比，最大的不同在哪儿？你又是怎么把这一套搭起来、让自己能高效出活、把东西真做出来的？
[8:29] Ryan
Totally. Um I like to say I'm in max Q right now for those of you that are are um you know SpaceX fans. It's like it the maximum pressure that I need to hire somebody. I'm doing all the work. And so yeah, what am I actually doing dayto-day? So I'll walk you through it. Um, I started off, you know, in uh a more specific agent like a cursor, um, like an AMP. Um, and and then I I I saw cloud happening. I'm like, "Okay, I'm really limited cuz I have to be at my machine and there's a lot of things I need to, you know, quick PRs like here's a bug, need to sort it out or what about this piece of data, need to, you know, and I was like, why why isn't cloud like I feel like I should be working on the cloud?" And I thought, what about Devon? Like Devon, you know, they've been doing cloud for like years now. It was kind of crap in the beginning, right? But maybe it's good now. And I tried it and I was like, "Wow, this this thing's really good. This is not
完全没错。我喜欢说我现在正处在 max Q——SpaceX 的粉丝应该懂，就是动压最大的那个时刻，压力大到我必须招人了。所有活都是我一个人在干。那我每天到底在干什么？我带你走一遍。我一开始用的是那种更具体的 agent，比如 Cursor，比如 Amp。然后我看到云端这条路起来了，我就想：这样我太受限了，因为我必须坐在自己机器前面。而我有一堆事要处理，比如快速提个 PR——这儿有个 bug 得修一下，或者这块数据怎么回事得看看——我当时就想，为什么不能在云上跑？我觉得我应该在云上干活才对。然后我想到：Devin 呢？Devin 他们做云端这套都好几年了。刚出来那会儿挺烂的，对吧？但说不定现在好用了。我试了一下，然后我就想：哇，这东西真的很好用。这不是
[9:28]
good.
好用。
[9:29] Ryan
It's really good." And you know, and I I thought, "Okay, I'm going to try it." And it was just blown away by how mature the cloud agent was.
是真的好用。然后我就想，行，那我试试。结果我完全被这个云端 agent 的成熟度给震住了。
[9:37] Ryan
And and if you're out there listening and you know, you are doing engineering, you're doing your work locally, you really need to open your eyes. Like I I think the future is pretty much 100%, you know, cloud agents. there there's going to be some local dev. But so I started looking at that and was really impressed. So now I went all in. So much so I spent at one point I was like spending $5,000 on dev and then it went up to $20,000 one month and I'm like whoa wait a minute I have a big Twitter account. I'm going to go talk to the cognition guys and tell them to give me free credits and they did. So [laughter] So anyway, I you know get 20 grand of credits per month in Devon. Great. But the point is I live in Devon. I live in the cloud, right? So all of my engineering almost I and it's Devon's right here on my screen. I've got probably, you know, 10 probably 10 15 threads and I bucket them in folders cuz I remember one of your shows, you know, it was either you or your guest was saying it's sort of like four or five concurrent tasks you can actually keep track of. And I agree with that. So what I use is folders.
所以如果你正在听这期节目，而你是做工程的、活儿还都在本地干，那你真得睁开眼看看了。我觉得未来基本上会是 100% 的云端 agent，当然还是会剩一部分本地开发。总之我开始研究这条路，印象非常深。于是我全押了。押到什么程度呢——有一阵子我一个月在 Devin 上花 5000 美金，后来一个月涨到 2 万美金，我就想，等一下，我 Twitter 账号粉丝不少啊，我去找 Cognition 的人，让他们给我免费额度。结果他们还真给了。所以我现在每个月能拿到 Devin 两万美金的额度，挺好。但重点是，我就活在 Devin 里，我活在云上，对吧？我几乎所有的工程活都在里面。Devin 现在就在我屏幕上，我大概同时开着十个、十来十五个 thread，然后我用文件夹把它们分类——因为我记得你有一期节目里，不是你就是你的嘉宾说过，人真正能同时盯住的大概也就四五个并行任务。我同意这个说法。所以我用的办法就是文件夹。
[10:47] Claire
Is there a way to show it? Um, I can show it. I'll show you one thing, which is this piece of paper. [laughter]
有办法展示一下吗？嗯，可以展示。我先给你看一样东西——就是这张纸。
[10:54] Claire
Perfect.
完美。
[10:55] Claire
And this is the reason we're doing where we say we're going to hire humans and we're just going to show you pieces of paper we scribble. [laughter]
这就是为什么我们嘴上说要招人类——然后要给你们看的，是我们自己涂涂画画的几张纸。
[11:05] Ryan
Okay. The re but there's a there's a I'm doing this for a reason. Okay. So, the reason why I have folders for threads and the reason why I have a piece of paper is because it the skill is managing throngs of agents. like that is what I do now. And to to keep it straight in your head, you basically have to bucket these things. And so I have uh one bucket um uh that I'll I'll try to show this to you. Let let's see how let's see how dangerous this is going to become everybody. Okay. So everybody, this is Devon. This is my real setup that I'm literally working on today. And you know I think one of my big messages for today is that all of us have to uplevel our ability to manage agents like that is our job right and so how do you do that? How do you effectively manage 10, 20, 100, 500 employees basically? Um,
好。原因是——不过这里有一个、有一个……我这么做是有原因的。是这样，我之所以给 thread 分文件夹、之所以要拿一张纸，是因为现在真正的核心技能就是管理一大群 agent，这就是我现在干的活。而要在脑子里理清楚，你基本上只能把这些东西分桶。所以我有一个桶……我试着给你看一下。大家看看这会不会玩脱。好，各位，这就是 Devin，这是我今天真正在用的、货真价实的工作台。我觉得我今天最想传达的一点是：我们所有人都得把「管理 agent」这个本事提上一个档次，这就是我们的工作，对吧？那具体要怎么做到？你要怎么高效地管 10 个、20 个、100 个、500 个「员工」？嗯，
[11:58] Ryan
this is the whole reason that organizations were built as pyramids,
这正是为什么组织当初都被搭成金字塔形，
[12:02] Ryan
right? Like humans can't scale and have a thousand directs unless you're like Jensen Mong then you can. But
对吧？人是没法无限扩张的，谁也带不了一千个直属下属——除非你是 Jensen Huang，那你确实可以。但是
[12:09] Ryan
so what I'm doing is figuring out how to manage, you know, throngs of agents. And in Devon, what what you're seeing is um is folders. And you'll notice um I've got bugs, right? I've got P 0, P1, P2. So, it's a priority zero. What What are the things that I I really need to move forward no matter how distracted I get today. And so, I'm going to show you again my piece of paper. Tada. Okay. So, I I have [laughter] two methods um for this. Don't laugh. It works. Um uh so the way I stay focused is number one, have a P 0 folder, but then you know I've got a Dell 52 here. I've got eight screens concurrently open, right? And so the way I stay focused is by also having this piece of paper which has weekly priorities on it.
所以我现在在琢磨的，就是怎么管好这一大群 agent。你在 Devin 里看到的，是一堆文件夹。你会注意到我有一个 Bugs，然后是 P0、P1、P2。P0 就是最高优先级——不管我今天被什么事打断，有哪些事是我无论如何都得往前推的。那我再给你看一次我那张纸。铛铛。好，为这件事我准备了两套办法（笑）。别笑，真管用。我保持专注的第一招，就是有一个 P0 文件夹；但你也看到了，我这儿摆着一台 Dell 52，同时开着八个屏幕。所以我保持专注的另一招，就是还有这张纸，上面写着这一周的优先事项。
[12:54] Claire
Oh,
哦，
[12:55] Ryan
okay. So, it's all about weekly what is important and and I'll and I'll cast my eye down here um at this beautiful UG Monk uh system I have, which I'm not being paid to say that. It's a fun little to-do system
好。所以核心就是：这一周什么最重要。我会时不时往下瞟一眼这套挺好看的 Ugmonk 系统——顺便说一句，没人付钱让我这么讲。就是个挺好玩的小待办系统，
[13:08] Ryan
and that just keeps me grounded. Okay, so let's dig into this a little bit. Bugs is like, okay, there's some stuff going on. I want to get it pretty fast. It's not necessarily P 0. Um, you know, let's look at a couple of those. You know, that's the one we're looking at here. Investigate red error retries. Yep.
它能让我一直踩在地上、不飘。好，我们再往里看一点。Bugs 就是：有些事正在发生，我想比较快地解决掉，但不一定到 P0 那个级别。我们挑几条看看。喏，我们现在看的就是这条——Investigate red error retries（排查报错重试）。对。
[13:25] Ryan
Right. So, Devon's cranking on that. Um, there's a couple other bugs in there.
对，Devin 正在啃这条。里面还有另外几个 bug。
[13:29] Ryan
P 0 is the big stuff. I'm not going to show you that, but that's just like these are the these are the the main things I need to move forward as the business. New features. Um, usually
P0 里放的是大事。这个我就不给你们看了，但基本上就是：作为一门生意，我必须往前推的那几件核心事情。然后是 New features（新功能）。通常来说，
[13:40] Ryan
P1 is like ah stuff that's important. Anyway, it goes down. Um, investors, by the way, tip. Oh my god. Like, if you're not using an agent to help draft your investor updates and help you keep track of the business, like clearly do that. But here's the key. Build a skill to do it. And this is probably obvious to everybody. But another thing I love about Devon is this idea has playbooks. and playbooks are are are sort of there's this weird mix of skills playbooks you know sort of uh thing going on but I've built an investor update skill basically which does a lot so this is most of where I do work and then you'll find like then that means I do a lot of work here like I'll be on my phone and this is the other reality is like unfortunately that means I can work like 21 hours a day And so you do
P1 是那些重要但没那么急的东西。总之优先级就这么一路往下排。对了，还有 Investors（投资人）这一档，顺便给个小建议——天哪，如果你还没用 agent 帮你起草投资人更新、帮你盯着公司的情况，那真的赶紧去做。但关键在这儿：把它做成一个 skill。这点可能对大家来说已经是常识了。我特别喜欢 Devin 的另一点，是它有 playbooks 这个东西。playbooks 嘛……现在 skills、playbooks 这些概念是有点混在一起、说不太清，但我基本上就做了一个「投资人更新」的 skill，它能干挺多活。我的大部分工作都是在这儿完成的，然后你会发现——这也意味着我在这上面花了非常多时间，比如我拿着手机也在干。这就是另一个现实：不幸的是，这意味着我一天能干上 21 个小时。所以你确实——
[14:35] Ryan
I do. And then you know and how do you deal with this? How do you deal with like the the the decision fatigue?
我确实是。那你呢，你怎么应付这个？你是怎么处理那种决策疲劳的？
[14:44] Claire
Yeah. So I want to reflect. I'm going to answer that question in just a second, but I want to reflect for folks what I'm seeing here because you and I are different AI managers
嗯。我想先说说我的观察。这个问题我马上回答，但我想先替大家梳理一下我在这儿看到的东西——因为你我是两种很不一样的「AI 管理者」。
[14:54] Claire
and every time you come on I'm like
每次你来上节目，我心里都在想：
[14:57] Claire
no. I'm like, what an organized what an organized individual. I am so impressed by you. You're like, here's my three steps to man my managing my agents. And I'm like, you don't just like float [laughter] in the tokens, see what comes out the other side. You're like,
不是吧。我就想：这人也太有条理了吧。我真的特别佩服你。你会说「我管 agent 就这三步」，而我心里想的是：你居然不是随手往 token 里一泡（笑），然后看那头能吐出什么来。你是那种——
[15:12] Claire
here are my like prioritized triaged folders for Devon. And I'm like, I just slack Devon. [laughter] Like he belongs to me. I just like a sentry link. WTF question fix
「这是我给 Devin 分好优先级、做过分诊的文件夹。」而我呢，我就是在 Slack 上直接找 Devin（笑），跟他是我的人似的。我就甩一个 Sentry 链接过去，配一句「这什么鬼？修一下。」
[15:27] Ryan
and that works. I mean there's nothing wrong.
那也管用啊。我是说，这没什么不对的。
[15:30] Claire
Yeah, it totally works. We're we're so different. But I do think you know something you said earlier I want people to hear which is like the job right now is to figure out how to manage agents. And the reason why I know both you and I are 27. So of course we're we're young. We're spring chicken infinite energy. The reason why I think folks of a certain level of experience are like ready to cook right now with AI is all those management skills that you have developed over your career,
对，完全管用。我们俩风格差太远了。但你刚才说的有一点我想让大家听进去：现在这份工作的核心，就是搞明白怎么管 agent。而我之所以——我知道咱俩都是 27 岁嘛，所以当然啦，我们年轻得很，精力无限。我觉得有一定经验积累的人现在特别容易在 AI 上做出成绩，原因就是你职业生涯里练出来的那些管理能力，
[16:00] Ryan
oh my god,
天哪，
[16:01] Claire
can be applied right here,
现在正好能用在这儿，
[16:04] Claire
right now.
就在当下。
[16:05] Claire
How do you structure? How do you team? How do you prioritize? How do you delegate? When you have a team of 500, how do you still stay anchored on like the three things that are most important for you to do? Yes.
怎么搭结构？怎么组队？怎么定优先级？怎么把活派出去？当你手下有 500 号人的时候，你怎么还能死死盯住对你最重要的那三件事？对。
[16:17] Claire
Where do you micromanage? Where do you let go? Like all that stuff still
哪里该盯细节，哪里该放手？所有这些东西，到今天依然
[16:22] Claire
pays off.
有回报。
[16:23] Ryan
Yeah. And it pays off. Yep.
对，是有回报的。没错。
[16:25] Ryan
I I think in multiples like um so another practical tip I want to talk about is I have um it's a playbook, but it could be a skill if you're not using Devon. It's called Watchdog. and and you know, we're starting to pick up a lot of family law firms now and they have a lot of cases, right? And and I'm literally doing customer success and engineering and sales, right? And so it's like, whoa, how do you keep track of actually what's going on in these cases? You know, where where are clients hitting UX bugs? Like where what oh my god, how do you keep track of it? So watchdog, what it does, and I this is multiple iterations, everybody. So what you want is like when that moment where you feel like I need to know what's going in the business right now. I feel stressed. I don't know what's happening. Uh this is when I run watchdog and and what it does is it basically it it has a skill for that that's repeated for each firm. So think of these as customers, right? Uh it and it goes into each account and says what's going on with this account like what was the activity since we last did watchdog? Where are the errors? Where are the sentry uh uh um errors? You know, where does all this happen?
而且我觉得是成倍放大的。我还想讲一个很实用的技巧：我有个东西——它在 Devin 里是个 playbook，但你要是不用 Devin，做成一个 skill 也一样。它叫 Watchdog。我们现在开始接到不少家事律所的客户，他们手上案子特别多。而我一个人同时在做客户成功、工程和销售。所以问题就来了：你到底怎么知道这些案子里正在发生什么？客户在哪儿撞上了 UX 的 bug？天哪，这些你怎么跟得过来？Watchdog 干的事是——先说好，这东西我迭代了很多轮才成型。你要的是这样一个时刻：你突然觉得「我现在必须知道公司里到底在发生什么，我很焦虑，我不清楚状况」——这时候我就跑一次 Watchdog。它本质上是有一个 skill，对每一家律所重复跑一遍。你把这些律所当成客户就行。它会进到每个账号里问：这个账号最近怎么样？从上次跑 Watchdog 到现在都有哪些动静？报错出在哪儿？Sentry 上的报错在哪儿？这些事都发生在什么地方？
[17:36] Ryan
But then but then you don't want just data. You want it to be transposed and filtered and sorted for you. So then it comes up with the top three things that are going wrong. And then either like, well, wait, have they been fixed? because we're shipping a lot of PRs. Like I'm shipping probably 40 PRs a day now. So, did it get fixed or is it currently being fixed or is there an open PR that never got merged? Right. And it just gives me all that and then I talk to that thread about okay, let's dig in a little bit. And that's how I manage the chaos.
但你要的不只是一堆数据。你要它帮你把数据转个向、筛一遍、排好序。所以它最后会给出「当前最糟的三件事」。然后还得接着问：等等，这些是不是已经修好了？因为我们在疯狂地发 PR——我现在大概一天发 40 个 PR。所以：这个到底修好了没？是正在修？还是有个 PR 开着一直没合进去？对吧。它把这些全都给我，然后我就在那个对话线程里接着聊：好，我们再往深挖一点。我就是这么管住这一团乱的。
[18:10] Claire
Yeah. I you know, you asked me how do I manage the the decision fatigue? And I feel like I start a business for every child I have. So like three kids, three businesses. I have a lot of decision fatigue.
嗯。你刚问我怎么处理决策疲劳。我感觉我是每生一个孩子就开一家公司，三个孩子、三家公司。所以我的决策疲劳可太多了。
[18:22] Claire
There's there's a couple things that that I do. I approach it very differently, which is I think about bucketing my time more than I think about bucketing the tasks, which is like
我有几个做法。我的思路跟你很不一样：比起给任务分类，我更多是在给自己的时间分块，就是说——
[18:32] Claire
there's time in the day that I do like very very specific tasks where I'm co-piloting with AI. There's other time where I'm just like
一天里有一段时间，我是在做那种非常非常具体的活儿，跟 AI 一起「双人驾驶」。还有另外一段时间，我是专门用来——
[18:39] Claire
I dedicate to spinning off everything or checking in. So I have to buck as a solo founder,
专门用来把各种任务派出去，或者挨个去查看进度。所以作为一个独立创业者，我必须——
[18:45] Claire
I have to bucket my own time
我必须给自己的时间分块。
[18:47] Claire
on things. The other thing that I use is I use decision fatigue as a priority prioritization heristic which is functionally if there is a feature that I'm working on that I let die on the vine.
……在这些事情上。我还有一招，是拿「决策疲劳」当优先级判断的启发式——说白了就是：如果有个我在做的功能，我一直放着让它烂在藤上、迟迟不推进，
[19:02] Claire
That is my [laughter]
那就是我的——（笑）
[19:04] Ryan
that's your symbol.
那就是你的信号。
[19:05] Claire
That's my symbol for
那就是我的信号，说明
[19:07] Claire
maybe this one is not the one I really care or believe in.
这个可能并不是我真正在乎、真正相信的那个。
[19:10] Claire
And so um I I sometimes think that people, you know, don't don't tell anybody. I know I'm a token maxer and you should let your agents run 24 hours a day and ship as much code as humanly possible and software everywhere. And I really don't ship more than I think the market
所以我有时候会觉得——这话别跟别人说啊——我知道我是个 token maxer，主张让 agent 一天 24 小时跑着、尽人类所能地多出代码、到处都是软件。但我其实真的没有超出我判断的市场
[19:29] Ryan
Yeah.
嗯。
[19:30] Claire
wants like I really try to res like constrain my output not quality not on bugs but like because I don't I don't think I get
需求量去发东西。我是真的会刻意约束自己的产出——不是在质量上、也不是在 bug 上，而是因为我不觉得我能
[19:40] Claire
multiples of quality off of multiples of output.
用翻倍的产出换来翻倍的质量。
[19:44] Claire
I totally agree. And so I would rather reserve that creative energy for something else or at least keep it inside till there is something there. And so I do um I run less overnight. I run like bugs and triage and all that kind of stuff. I also work with an engineer so we kind of pingpong
我完全同意。所以我宁愿把那股创造力留给别的事，或者至少先攥在手里，等真有点东西了再拿出来。所以我晚上跑得比较少，跑的多是修 bug、做分诊这类活儿。我还跟一位工程师搭档，所以我们俩会来回接力，
[20:01] Claire
overnight but I don't I don't do the sort of like seek out ideas, solve them and always ship them because I just don't one it it puts too much on me and two I don't think it actually adds value.
通宵跑。但我不会去做那种「自己找想法、自己解决、然后一律发出去」的事，因为我不干这个：一是它给我自己的负担太重，二是我不觉得它真的创造了价值。
[20:13] Ryan
I agree. I I actually think we're nowhere near any frontier model having the intelligence to know what to ship.
我同意。我其实觉得，要说有哪个前沿模型聪明到知道「该发什么」，我们离那还差得十万八千里。
[20:21] Claire
I I agree.
我同意。
[20:22] Ryan
It's it's like a million miles. I mean
那真是差了一百万英里。我是说——
[20:24] Claire
I agree.
我同意。
[20:25] Claire
So So yeah, I think you know I think we all wanted these sort of like automatic improvement loops to and they just don't work with product.
所以，对，我觉得我们当初都想要那种自动改进的闭环，但它在产品这件事上就是行不通。
[20:32] Ryan
They're kind of a joke.
那玩意儿有点像个笑话。
[20:33] Claire
And and I and I tell people it's also because all of this intelligence has not magically created markets where there were none before
而且我也跟别人说，这还因为：这么多智能并没有凭空变出原本不存在的市场，
[20:41] Claire
other than AI markets, right? It's like it's not like all of a sudden there are these like buyers that have shown up for solutions that somebody Yeah. for stuff.
除了 AI 本身那些市场之外，对吧？又不是说突然之间冒出来一堆买家，来买谁谁做出来的那些方案——是啊——来买那些东西。
[20:51] Claire
And so I do think there's this like mismatch between the quantity of output of code and
所以我确实觉得，代码产出的数量，跟
[20:59] Claire
like commercializable product.
真能商业化的产品之间，存在这种错配。
[21:02] Ryan
Amen. So I want to double click on that. I think what what what is happening is people are not getting out of their chair enough and actually talking to real people. So, so you know the way I figured out PMF for Untangle was, you know, I went outbound with an email, I did a a Google meet with a potential customer and I talked to them. Boom. That was step one. But then step two is like as soon as we landed them as a customer, I was like, "Can I come to your office?" And I just want to sit and I want to talk and I want to and this is all like obvious product stuff, but I think because we can ship so much and build so much digital product, we're we're not getting out of our chairs enough.
说得太对了。我想就这点再往下挖一层。我觉得现在的问题是，大家不够常从椅子上站起来、去跟真实的人聊。比如我当初怎么给 Untangle 摸到 PMF 的：我发了一封外呼邮件，跟一个潜在客户开了个 Google Meet，跟他们聊。搞定，这是第一步。但第二步是，一把他们签成客户，我马上就问：「我能去你们办公室吗？」我就是想坐在那儿、想跟他们聊、想……这些都是很基本的产品常识，但我觉得正因为我们现在能做出、能发出那么多数字产品，我们反而越来越不肯离开椅子了。
[21:43] Ryan
Um, so
嗯，所以
[21:44] Claire
I totally agree.
我完全同意。
[21:45] Ryan
Go talk to humans for real.
真的去跟活人聊。
[21:47] Claire
Talk to humans. Yep.
去跟人聊，没错。
[21:48] Ryan
And and build something neat.
然后做点漂亮的东西出来。
[21:50] Claire
This episode is brought to you by Jira by Atlassian. The teamwork graph in Jira delivers 44% more accurate agent results with 48% less token usage. That's a huge difference when working with AI coding agents like Claude, Cursor, Codeex, or Copilot. The hardest part of shipping with AI isn't [music] the code, it's the context. What's the right ticket? What did the spec say? What got decided in Slack? The teamwork graph pulls all of that from across your entire stack from Jira and Confluence to GitHub and feeds it directly to your agents before they write a single line. You assign the work, the agent [music] gets everything it needs and a PR surfaces when it's ready. No digging, no context switching, no broken flow. Same team, smarter agents. Try them free at jira.dev. That's j I R A.dev. The other thing I want to touch on because I think you breezed past it, but it has been a gamecher for me in terms of using background agents and in particular Devon, which is I think people have still bucketed these cloud coding agents as coding agents, which is like their [clears throat] only job
本期节目由 Atlassian 旗下的 Jira 赞助。Jira 的 teamwork graph 能让 agent 的结果准确率提升 44%，同时少用 48% 的 token。跟 Claude、Cursor、Codex 或 Copilot 这类 AI 编码 agent 一起干活时，这个差距是巨大的。用 AI 出活儿，最难的（音乐）不是代码，是上下文：该看哪张工单？spec 里到底怎么写的？Slack 里最后定了什么？teamwork graph 会把这些东西从你整个技术栈里捞出来——从 Jira、Confluence 一直到 GitHub——在 agent 写下第一行代码之前就直接喂给它。你把活儿派下去，agent（音乐）就拿到了它需要的一切，做完了自然会冒出一个 PR。不用翻找，不用来回切上下文，心流也不会断。同一支团队，更聪明的 agent。去 jira.dev 免费试用，就是 j-i-r-a.dev。另外还有一件事我想聊，因为我觉得你刚才一带而过了，但它对我来说是个 game changer，就是用后台 agent，尤其是 Devin。我觉得大家到现在还是把这些云端编码 agent 归类成「编码 agent」，好像它们唯一的活儿（清嗓）
[23:04] Claire
is to write code for a product or software product. And in fact, Devon runs Deal Desk for me.
就是给某个产品、某个软件产品写代码。但实际上，Devin 在帮我跑 Deal Desk（报价台）。
[23:13] Claire
So Devon's doing quoting, custom quoting for me.
也就是说，Devin 在帮我做报价，做定制报价。
[23:17] Ryan
Yeah.
对。
[23:17] Claire
Like setting out operations, um, documentation, customer triage, like
比如把运营的事安排起来、写文档、客服工单分流，就这类——
[23:26] Claire
I I always think, you know, what would you do with somebody who both knew your code base and could write code to solve anything in your business? That's how I use these these background agents. And I do think that um people really don't ask enough of their back background agents and especially of ones that are very good at coding. I think this is why it's so interesting. Cursor just yesterday came out or cursor or Grock whoever who whoever they are at these came out with Grockbot and I was like you know what the cursor chat experience as a as a background agent was a great one. It was like one of our first first episodes. And so I'm not surprised to see these like coding platforms start to generalize into business agents in general. And I also feel like more than engineers need to get hands-on with with background agents and cloud cloud agents because there's so much that they can do for your business pretty autonomously.
我老是会想：如果你身边有这么一个人，他既懂你的代码库，又能写代码解决你业务里的任何问题，你会让他去干什么？我就是这么用这些 background agent 的。而且我真觉得，大家对自己的 background agent 要求太少了，尤其是那些编码能力特别强的 agent。这也是我觉得特别有意思的一点。Cursor 昨天刚发了个东西——是 Cursor 还是 Grok 来着，反正不管是他们哪家——发了个 Grokbot。我当时就想，Cursor 那种把对话式体验做成 background agent 的路子真的挺好的，我们最早那几期节目里就聊过。所以看到这些编码平台开始往通用的业务 agent 上扩，我一点都不意外。我还觉得，不该只有工程师去上手 background agent 和云端 agent，因为它们能相当自主地帮你的业务干太多事了。
[24:26] Ryan
100%. Um but let's talk about though why probably both of us are still using local agents uh for some things. So it's so right now as well I have codeex running um
百分之百同意。不过我们也聊聊，为什么我俩到现在有些活儿还是在用本地的 agent。比如我现在也开着 Codex，
[24:38] Ryan
and I have and I have it using goal
而且我让它用 goal，
[24:41] Ryan
and I find that in particular like the low latency plus the browser on my machine pairing like plus like basically infinite tokens you know from from
我发现特别是这种组合——延迟低，再加上能跟我本机上的浏览器配合，再加上基本上是无限的 token，来自——
[24:53] Ryan
from OpenAI. Um
——来自 OpenAI 的。
[24:55] Ryan
yeah cool. I'll take the subsidy. Um, that, you know, plus goal really, really is good for when you have this kind of I've got a feature, it's a lot of front-end UI. Yep.
对，挺好，这个补贴我就收下了。就是说，这些再加上 goal，真的特别适合那种场景：我手上有个功能要做，前端 UI 的活儿特别多。对。
[25:06] Ryan
I just need you to grind for a lot and I need to be here and kind of hold your hand. Y,
我就是需要你埋头猛干一大堆，同时我得守在旁边，多少要手把手带着你。嗯，
[25:10] Ryan
I find Codex unmatched for that.
干这种活儿，我觉得 Codex 无可匹敌。
[25:12] Claire
Yeah. So, if I were to because what we said we would do is compare and contrast our solo founder AI setup. So I would say I am personally spending my time in codecs
对。那如果让我——因为我们说好了要把我们俩作为独立创始人的 AI 配置对比一下——我会说，我个人的时间主要花在 Codex 上，
[25:23] Claire
when I'm pair programming or or managing a product. It's it falls into a couple categories. It's going to be like large feature initiatives that span front end and backend that both like again need like a lot of code written and you have to grind on it and needs that like intervention every now and then to narrow it to the experience that you want. um a lot of like long running refactors and hectets stuff like again I think the latency is unmatched it's goal is really really effective sub agents are really really effective um and then I use codeex a lot for ver I think it's very good at verification and so a combination of goal plus a CLI tool inside your codebase that can verify core flows or I use goal and browser use to do front-end verification a lot. So I say write user stories, use Chrome, pull up the preview branch in Chrome, go through it as a user, list which user stories pass, identify bugs, fix them until you're good. Those are kind of what I'm using codecs for. When I'm using background agents, it's a lot of like autofixes on bugs. um oneoff like when I'm on calls with with customers and they say, "Oh, do you have this MCP?" I'm like, "I do now." Um so oneoff things and then sort of um operations as you said, like I use a lot of agents for for operations and then you know I do have a couple like purpose-built agents in the stack. So, I use Bugbot a lot for PR review and then I just built this thing delightfully called Mergemmy that that does risk assessment on my PRs and auto approves lowrisk ones. Um, because I am also shipping a lot of PRs
就是在结对编程、或者在管一个产品的时候。它大致分几类。一类是那种横跨前端和后端的大功能：同样地，需要写很多代码，得死磕，而且时不时需要你介入，把它收敛到你想要的那个体验上。还有很多是长周期的重构、架构层面的活儿——还是那句话，延迟这块我觉得没人比得过，它的 goal 非常非常好用，sub agent 也非常非常好用。另外我还大量用 Codex 做验——我觉得它做验证特别在行。所以就是 goal 加上一个能在你代码库里跑的 CLI 工具，用来验证核心流程；前端验证我也经常用 goal 配 browser use。我会这么说：写用户故事，用 Chrome，在 Chrome 里把 preview 分支调出来，以用户的身份走一遍，列出哪些用户故事通过了，找出 bug，一直修到没问题为止。这些差不多就是我拿 Codex 在干的事。而我用 background agent 的时候，很多是自动修 bug；还有那种一次性的活儿——比如我在跟客户开会，他们问「你们有这个 MCP 吗？」我就说「现在有了。」所以是这种临时需求。再就是你刚说的运营那块，我用很多 agent 做运营。另外我这套里确实还有几个专门定制的 agent：PR review 我大量用 Bugbot，然后我最近还做了个东西，名字起得挺有意思，叫 Merge Mommy，它会给我的 PR 做风险评估，低风险的自动批准。因为我每天发的 PR 也很多，
[27:14] Claire
and so that's
所以那就是——
[27:15] Ryan
and is that
那这个是不是——
[27:16] Ryan
tell me about Mer tell me about merge mommy like that's and then I'll tell you about my land PR skill. I want to compare and contrast.
跟我讲讲 Mer——跟我讲讲 Merge Mommy，因为这个……然后我再跟你讲我的 land PR skill。我想对比一下。
[27:23] Claire
Yeah, so I did an episode on this a week or two ago. It's an EVE agent, so it's deployed on Burcell. And whenever a PR is opened and then passes all its CI checks, including Bugbot, Merge Mommy gets triggered. And then what she looks at is she scores the PR on like five aspects of risk, like cool,
好。这个我一两周前专门做过一期节目。它是个 EVE agent，部署在 Vercel 上。只要有 PR 被打开、并且通过了全部 CI 检查（包括 Bugbot），Merge Mommy 就会被触发。然后她做的事情是：从五个风险维度给这个 PR 打分，比如说——
[27:43] Claire
how big a blast radius, security, whatever it is. And then if it um gets scored as a low, Merge Mommy stamps it with an approval and we can merge without a human approval.
——影响半径有多大、安全性之类的。然后如果打分结果是低风险，Merge Mommy 就直接盖个批准章，我们不用人来批就能合。
[27:56] Ryan
Cool.
酷。
[27:56] Claire
But if it's medium or high, then it pings us in Slack and says it pings um either my colleague or I or somebody else in Slack and says this one's ready for you to review because for our sock 2 compliance and all our controls, we require
但如果是中风险或者高风险，它就会在 Slack 里 @ 我们——@ 我同事，或者我，或者 Slack 里的其他人——说这个 PR 该你来 review 了。因为要满足我们的 SOC 2 合规和那一整套控制要求，我们必须有
[28:10] Claire
kind of like alternative review.
那种由另一个人来复核的 review。
[28:12] Ryan
Um and it just allows us to like get through our PRQ very effectively. Yeah. Yeah, first of all, annoyed I missed that episode. That sounds amazing. I'm going to go watch it. Um, but number two, interesting. So I have something called land PR where basically it's it's a playbook in Devon but you can think of it as a skill and it's bas usually it's when the agent comes back and says okay the PR is done you know and and then uh I trigger it and then what it does is it basically makes sure that number one it runs a fresh Devon review on the PR which and Devon review is like an actual product inside of Devon which is another reason why I think people should just pick I'll I'll sort rant about that later, but the code factory like it should have a good a good built-in review that is a real review product, right? So, it triggers a review and then it basically allows up to two loops on that, right? you know, okay, all right, find all the all the bugs, check that all the comments are are are resolved, and then and then there'll be some more that come up, do it again, does it twice, and then then the thing is is it records a a video walkthrough. Um, and and this I find Devon is just really good at like, so what it does is it records in the browser and it actually kind of narrates over it. It's got captions about it's doing and it's got this cool little little test, you know, list of tests it's doing and whether they're red or green and you get all this out of the box. Like I tried to build my own code factory and get all this stuff and I was like I don't have time to do this and so the land PR basically it has to show me that video and then I then I say video approved land it and then it will merge.
这样我们清 PR 队列的效率就特别高。嗯。嗯，首先，我居然错过了那期节目，好气，听起来太赞了，我等会儿就去看。第二，很有意思。我这边有个东西叫 land PR，它本质上是 Devin 里的一个 playbook，但你可以把它理解成一个 skill。一般是 agent 回来跟我说「好了，PR 做完了」的时候，我就触发它。它做的事是：第一，在这个 PR 上重新跑一遍 Devin review——Devin review 是 Devin 内置的一个真正的产品，这也是我为什么觉得大家就该选定一个平台，这个我等下再专门吐槽——就是说你的 code factory，它应该自带一套好用的 review，而且得是个真正的 review 产品，对吧？所以它先触发一次 review，然后最多允许在这上面循环两轮：好，把所有 bug 找出来，确认所有评论都处理掉了，然后又会冒出来一些新的，再来一遍，一共两轮。接着关键的是，它会录一段视频走查。这块我觉得 Devin 做得是真的好——它会在浏览器里录屏，而且还会配上解说，画面上有字幕说明它正在做什么，还有个很酷的小测试清单，一条条列着是红还是绿，这些全都是开箱即用的。我以前也试过自己搭一个 code factory 把这些都做出来，然后我就想，我哪有这个时间。所以 land PR 的核心就是：它必须把那段视频给我看，我看完说「视频通过，合了吧」，它才会去 merge。
[29:56] Ryan
Um so that really works for me. It's kind of that I think this is where CI/CD and DevOps like now you have to be even better at it like and really care about it and and we use merge cues and GitHub as well. I'm loving merge cues. Um so couple of the nitty-gritty this is what I'm actually doing stuff.
所以这套对我特别管用。我觉得这也正是为什么现在 CI/CD 和 DevOps 你得做得更好、而且要真的重视它。我们在 GitHub 上也用 merge queue，我现在特别喜欢 merge queue。所以说几个细节吧，这些都是我实际在干的事。
[30:13] Claire
Have you Okay, you all know this is how you know. Not not sponsored by Cognition, but you all are welcome. Um we love you. [laughter]
你有没有——好吧，你们都听出来了，就是这么听出来的。Cognition 并没有赞助我们啊，不过不用谢。我们爱你们。（笑）
[30:20] Ryan
Yeah, come on.
对，来吧。
[30:21] Claire
Um I can't get Devon video verification to work. I can occasionally get it to screenshot
我这边 Devin 的视频验证一直跑不通。偶尔能让它截个图，
[30:30] Ryan
really
真的假的
[30:30] Claire
and like on front end stuff it taps my um my credits for the session and then just falls over. I don't know if it's probably a problem. It's probably a mi problem. All AI things are probably a mi problem.
而且在前端这块，它会把我这个 session 的额度全吃光，然后就直接趴窝了。我也不知道，这可能是个问题，也大概率是我自己的问题。AI 上的所有问题大概都是我自己的问题。
[30:43] Ryan
No,
不会吧，
[30:44] Claire
but I have not been able but I actually I do want to show this for folks that um want to know exactly what it looks like because I found an example of it. So I'm claiming that it doesn't work but then I'm going to make it work. Hold on. Devon will actually go ahead and test it and record a little video of it using and you can see these little captions down here.
但我一直没能——不过我确实想给大家看看它到底长什么样，因为我找到了一个例子。我刚才还说它不好使，现在我又要让它当场好使给你们看。稍等。Devin 会真的去把它测一遍，还会录一小段使用过程的视频，你能看到下面这些小字幕。
[31:09] Claire
Uh it's exposing it's its reasoning. It's super cool. It's very useful. I can just only get it to work
它把自己的推理过程都露出来了，超酷，也非常有用。我就是只能偶尔让它跑通——
[31:17] Ryan
so a couple interesting comments I have in that. Number one, interesting that you're in Slack. Um
我对这段有几点挺有意思的观察。第一，有意思的是，你是在 Slack 里干活的。嗯——
[31:21] Claire
Oh, I I'm never in Devon UI. Never.
哦，我从来不进 Devin 的界面。从来不进。
[31:24] Ryan
Okay. So, interestingly, I'm never in Slack. Um, I am trying to cut Slack completely out of Untangle and and and the way the company is going to exist. Um, and I was laughing because you have so many sidebar, you know, um, notifications and exclamation points and and we were saying, you know, you're seeing the Devon preview in Slack and and I have a a very kind of perfect experience with the video playback testing of Devon and that's cuz they use the web app. And so here's what I'm thinking, and again, this is probably going to change, but as we hire people, I'm actually going to require like multi-user uh interaction to happen through Devon. So, we talk in threads about what is going on. That's thing one cuz cuz then you're not trying to always figure out where's the context for this thing, you know? Is it in your agent? Is it in Slack? Is it in Google Docs? Like where is it? So, I'm going to try to do that. Now, obviously that doesn't work for DMs and channels and all that stuff, but I also think and Jack Dorsey, I think, was saying this, you know, they launched Buzz. It's kind of interesting. I think he banned DMs or somebody banned DMs in their company. Um, and I think, you know, as we scale, I'm going to do the same thing. Like, I don't want any one-on-one chats. Like, if we're having private conversations, I'll have to call you. Um, so that's why I I just think Slack is not built for the AI age.
好。有意思的是，我几乎从不进 Slack。我正打算把 Slack 从 Untangle 里彻底砍掉，连带这家公司以后的运作方式也照这个来。刚才我笑，是因为你侧边栏里那么多通知、那么多小红点感叹号。我们刚才说到，你是在 Slack 里看 Devin 的预览，而我这边看 Devin 那个视频回放测试，体验几乎完美，那是因为他们用的是 web app。所以我的想法是这样——当然，这套以后大概率还会变——等我们开始招人，我打算要求多人之间的协作都发生在 Devin 里：大家就在 thread 里讨论当下在做什么。这是第一点，因为这样你就不用老去找「这件事的上下文在哪儿」：在你的 agent 里？在 Slack 里？在 Google Docs 里？到底在哪儿？所以我打算这么试。当然，DM、频道那一套显然是没法这么搞的，但我还觉得——Jack Dorsey 好像也说过类似的话，他们不是发布了 Buzz 嘛，挺有意思的——我记得是他、或者他们公司有人，把 DM 直接禁掉了。我觉得等我们规模起来，我也会做同样的事：我不想要任何一对一私聊。真要私下谈，那我就得给你打电话。所以我才说，Slack 不是为 AI 时代设计的。
[32:52] Claire
Well, let me let me tell you why I love Slack.
那我来说说我为什么爱 Slack。
[32:56] Ryan
Okay.
好。
[32:57] Ryan
I like to convince me.
来，说服我。
[32:59] Claire
I like to person first. I like to personify my agents. I like to as a solo founder, [laughter]
我喜欢先把它当个人看。我喜欢把我的 agent 拟人化。作为一个 solo founder，我喜欢……（笑）
[33:04] Claire
I need my colleagues to be fun.
我需要我的同事有意思一点。
[33:06] Ryan
Fair.
有道理。
[33:07] Ryan
Fair.
有道理。
[33:08] Ryan
I'm just boring and meme.
我就是那种又无聊又刻薄的人。
[33:09] Claire
Um, it's it's marginally more fun in Slack to use Dev than in the web app. Although I they're interchangeable actually. You're just seeing a reflection of the web app.
嗯，在 Slack 里用 Devin，确实比在 web app 里稍微好玩那么一点点。不过其实两边完全可以互换，你在 Slack 里看到的就是 web app 的一个投影。
[33:18] Claire
The the reason why I think I had been so um committed to Devon in Slack is actually not because I think it's a better experience. It's that very early on deploying background agents inside companies.
我之所以这么死心塌地在 Slack 里用 Devin，其实并不是因为我觉得那个体验更好。而是因为在公司内部铺开后台 agent 的最早期——
[33:33] Claire
This is where we're going to agree to agree inside companies. So when I was at Launch Darkly and we were first deploying background agents in particular Devon, it was such an effective change agent move to put these background agents in public Slack threads.
在「公司内部」这件事上，咱俩肯定是一致的。我在 LaunchDarkly 的时候，我们最早部署后台 agent、尤其是 Devin，把这些后台 agent 放进公开的 Slack thread 里，作为推动变革的一步棋，效果实在太好了。
[33:47] Ryan
Interesting.
有意思。
[33:48] Claire
And so like it was one of the places actually it was the only place I had experienced up to that point where you could do real time coding work in public.
那算是其中一个地方——其实是我在那之前唯一见过的地方——能让写代码这件事实时地、公开地发生。
[33:58] Ryan
Cool.
酷。
[33:58] Claire
And I think that changes people's perception of oh what is AI capable of doing? How do I work with AI better? Um, and then I do think, you know, in particular, Devon is pretty good at multiplayer um chat. It's one of the few ones that like you don't trip over your colleagues when you're both both chatting in a thread
我觉得这会改变人们的认知：哦，原来 AI 能干这些？我该怎么跟 AI 配合得更好？另外我确实觉得，Devin 在多人聊天这块做得相当好。它是少数几个能做到「你和同事同时在一个 thread 里说话也不会互相绊住」的。
[34:19] Ryan
in Slack.
在 Slack 里。
[34:20] Claire
But I think it's yeah uh in both in Slack and in in in the web app.
不过我觉得，嗯，在 Slack 里和在 web app 里都行。
[34:24] Claire
And I think um it's just I'm a creature of habit having told everybody to shove their Devon threads in public channels. Yeah. um that that's why I'm using it. Okay.
还有就是，我这人习惯成自然吧——当初我逼着所有人把 Devin 的 thread 都塞进公开频道，对，所以我自己也就一直这么用了。好。
[34:35] Ryan
You know, I'm
你知道，我——
[34:36] Ryan
I do think that chat's not going away because I think that's how humans like to
我确实觉得 chat 不会消失，因为人本来就喜欢这么
[34:44] Ryan
work like communicate. So, it's hard to feel like chat's going away going away. But I do think more and more people if they don't want to get rid of Slack, want to get rid of Slack culture, which is like DM culture, private channel culture, like I don't want any of that crap.
工作、这么沟通。所以很难觉得 chat 会彻底消失。但我确实觉得，越来越多的人就算不想干掉 Slack，也想干掉 Slack 文化——就是那种 DM 文化、私密频道文化。那些破玩意儿我一样都不想要。
[35:03] Claire
Needless needless conversation culture that doesn't result.
就是那种毫无必要、又什么都产不出来的对话文化。
[35:06] Ryan
It doesn't do anything. Yeah. And so almost like every thread in Slack should result in work.
什么都成不了。对。所以说，Slack 里几乎每一条 thread 都应该最终落成实际的活儿。
[35:13] Claire
Yep. But it doesn't
对。但它并不会——
[35:14] Ryan
feels better. It doesn't.
感觉上是舒服了。但其实并没有。
[35:15] Ryan
It devolves into just you know and I love humans, trust me. But I I don't listen, you know, I've I did a stint at Intel where I was an employee and I learned a lot about being what it's like to be an employee,
最后就退化成——你懂的。相信我，我是很喜欢人的。但我不……我在 Intel 待过一阵子，是以员工的身份，那段经历让我特别真切地明白了当一个打工人是什么滋味，
[35:28] Ryan
you know, and I just realized you spend a lot of time talking about stuff that doesn't matter. you're not empowered to do anything and it's like, you know, what's the point of that? So, and and you're the agent is always going to need the context of whatever you're talking about. So, so have the have the conversation there. So, before we leave, I know you have a hard stop. Um, I do want to talk about the third plane uh that I have, which which is a little bit of cloud code. So, I I Yeah. So, I'm number one, it's like it feels like infinite tokens. Um, but I just can't get around how good claw design is. Now, I don't mean in coming up with new design. I think it's bad. But I think for carrying forward, you know, a design system, it it it does something magical that Codex is hilariously bad at. Devon is terrible at at design and so I always have to go back to claw design. So, am I missing something or do you do something magical design that's working for you? You know, this is a this is a great question. I was telling you before this, I'm launching a new business. We've got a new brand, new design, and so I went through this process very recently. So, I can tell you
你知道吗，我就发现，你会花大量时间去聊那些根本不重要的事，而你又没被授权去做任何决定，那这有什么意义呢？所以说，反正 agent 早晚都需要你聊的这些内容作为上下文，那干脆就在它那儿聊。那在结束之前——我知道你时间卡得死——我还想聊聊我的第三个平面，也就是会用到一点 Claude Code。首先，它给我的感觉像是 token 无限管够。但更关键的是，Claude 的设计能力好到我实在绕不过去。我不是说让它想出全新的设计，那方面我觉得它挺差的。但在把一套设计系统延续下去这件事上，它能做出某种很神奇的东西，而 Codex 在这块差得有点好笑，Devin 的设计也很糟糕，所以我总得回头去用 Claude 做设计。那，是我漏掉了什么，还是说你在设计上有什么特别神奇、特别管用的做法？——这个问题问得太好了。录之前我跟你说过，我在做一门新生意，有新品牌、新设计，所以这套流程我最近刚刚走过一遍。我可以跟你讲讲
[36:35] Ryan
what I did. I think
……我具体是怎么做的。我觉得
[36:38] Ryan
Claw Designs create a design system
Claude Design 做设计系统这件事，
[36:41] Ryan
is like unmatched right now. Okay. It's it does it does a great job of decomposing
现在真的是没有对手。好。它非常擅长把
[36:47] Ryan
either an example site or a Figma or something into relative non-slop. design tokens you can reuse across across um
不管是一个参考网站，还是一份 Figma 稿之类的东西，拆解成相对不那么粗糙、没有 AI 味的 design token，让你可以复用到
[37:00] Ryan
across a property. Now that being said, my most recent processes I went to cloud design. I created
整条产品线里。不过话说回来，我最近这套流程是：我先去 Claude Design，做出了
[37:06] Ryan
the the kind of like design and the the design file in in Claw Design. I downloaded that, gave it to Codeex
那种设计稿、那个设计文件，就在 Claude Design 里做的。然后我把它下载下来，扔给 Codex，
[37:16] Ryan
and said, "Take this plus what's in the Figma and actually build me a technical design system."
跟它说：「把这个，再加上 Figma 里的东西，给我真正搭一套技术层面的设计系统出来。」
[37:23] Ryan
Oh. And it built all the shared components. It made this nice like interlinking in our mono repo.
哦。它就把所有共享组件全都建好了，还在我们的 monorepo 里做了一套很漂亮的互相引用关系。
[37:30] Claire
Oh wow.
哇哦。
[37:31] Ryan
Um and then I I also tested with codeex using the Figma plugin to then go pull all those designs tokens back into Figma into components that I could use.
然后我还试了一下，用 Codex 配 Figma 插件，把那些 design token 再反向拉回 Figma，变成我能直接拿来用的组件。
[37:45] Claire
Um
嗯——
[37:46] Claire
so once you have the design MD imported it to imported it to codeex. Yeah, that's exactly exactly what I did. I mean, my problem with Claude infinite tokens, very intelligent, my blood boils reading Claude output. It like actually raises my blood pressure. It makes me so. [laughter]
所以你有了 design.md 之后，就把它导进 Codex。（Ryan：对，我就是这么干的。）我是这样，我对 Claude 的意见是——token 感觉无限、也确实聪明，但我一读 Claude 的输出就火冒三丈。它是真的会让我血压升高，让我特别……（笑）
[38:08] Ryan
So, I the the one podcast I listened about the like the the the opinion on Opus 5 was yours
所以说，关于 Opus 5 的看法，我听过的唯一一期播客就是你那期，
[38:15] Ryan
cuz I I love your take and I couldn't agree more. I'm so sick of Opus 5. like it it's just so frustrating. It it's terrifying in one way cuz it's like oh my god it's way worse than we thought
因为我太喜欢你的观点了，简直不能再同意。我已经受够 Opus 5 了，真的太让人抓狂。它某种意义上还挺吓人的，因为那感觉就是：天哪，情况比我们以为的糟糕得多——
[38:28] Ryan
you know about everything
你懂的，方方面面都是。
[38:30] Ryan
and so you're like okay you know which is actually what's wrong here is and I so it's like super smart but s I really am sick of Opus 5. Yeah.
然后你会想，行吧，它其实真的能指出问题到底出在哪儿。所以它是真的超级聪明，但我真的受够 Opus 5 了。嗯。
[38:40] Ryan
Yeah. But I do think I the the applications in some like very specific deep spikes cyber security design like kind of unbeatable on the on the execution side still include but
嗯。不过我确实觉得，在一些非常具体的深度攻坚场景里——网络安全、设计这一类——它在执行层面基本还是无敌的，这点还是得算上，但是
[38:52] Ryan
um I am I just the the harness experience of codeex is just a lot more effective and enjoyable for me.
对我来说，Codex 那套 harness 的使用体验就是高效得多，也用得舒服得多。
[39:00] Claire
It's so good. So, last question I have before we go is like
太好用了。那在我们结束之前，我最后一个问题是：
[39:03] Claire
we can both answer, you go first. Like, what's your current actual um EA? Like, who's your actual sidekick that does all of your agent work as an executive assistant?
咱俩都答，你先来。就是，你现在真正的 EA 是谁？也就是说，那个替你打理所有 agent 活儿、相当于你行政助理的搭子，到底是谁？
[39:14] Ryan
Paulie the OpenClaw.
Paulie，我那个 OpenClaw。
[39:16] Claire
Interesting. Okay. You're on OpenClaw.
有意思。好。你用的是 OpenClaw。
[39:18] Ryan
Still live in OpenClaw.
我还是天天泡在 OpenClaw 里。
[39:19] Claire
Okay.
好。
[39:20] Ryan
Just committed there. [laughter]
就认准它了。（笑）
[39:22] Claire
Yeah. I love OpenClaw. I love Pete. I love the team. I love it. I just moved off. Um I was like, screw it. I'm just going to run cloud code um on my Mac Mini. Um and I actually want to switch to codecs cuz
嗯。我很喜欢 OpenClaw，喜欢 Pete，喜欢那个团队，是真喜欢。我只是换掉了。我当时想，管它呢，我就在自己的 Mac Mini 上跑 Claude Code 算了。而且我其实还想换到 Codex，因为
[39:35] Claire
I just love love love the UI of codecs. It there's so it's the it's wonderful the whole thing. So
我实在是太爱太爱 Codex 的界面了。整个东西都特别棒。所以
[39:42] Ryan
yep. Yeah. No, I use I still poly the open claw is still still running
对对。嗯。不过我还是在用——Paulie，那个 OpenClaw，现在也还在跑着，
[39:48] Ryan
but I do now have my emails constantly open in the browser in Codeex and Codex functionally runs my email. It it it it ran through 2,000 unread emails in my Gmail the other day and just changed my life.
但我现在会一直在 Codex 里开着浏览器挂着我的邮箱，等于是 Codex 在替我管邮件。前几天它一口气把我 Gmail 里 2000 封未读邮件全过了一遍，直接改变了我的人生。
[40:04] Ryan
This is this has been a game changer. Um so like you I've hired a lot of people in my life like
这真的是个颠覆性的变化。所以说，跟你一样，我这辈子雇过很多人，
[40:10] Ryan
hundreds and hundreds of people and and let go lots of people and and all this kind of stuff. So um I the way I'm running my hiring process right now really love it and we'll see if it works out. But but I think I think this is interesting. Um, so what I decided to do is I I basically on Twitter I was like, "Okay, we're hiring our first engineer. You got to be like super duper duper AI forward, you know, of course, and I don't really care where you live. And all I want you to do is record a video full screen, everything, your your whole desktop of you building a new feature for an app that already exists. And I just want to see the whole thing. Um, and I don't want to talk to you. I don't want to have a meeting with you. I don't want to get to know if I like you. I just want to see how good of an agent manager you are. And uh the videos I'm getting are fascinating and very revealing about whether I would want to work with this person. And I'm liking it. So, I've got, you know, a number of people that are probably going to move to phase two where then what I'm do is say, okay, give them access to Devon and actually build a real thing and then but record a video of that and then basically at the end I'll do some sort of quick, well, we should meet, you know, and chat and then I'm going to hire them and then I'm just going to watch their agent output and that's how I'm going to manage them. I mean, well, one other thing is if they work with Devon, you get the replay video anyway of how the Devon computer works. So, save you saved you a step.
……成百上千号人，也裁过很多人，诸如此类。所以呢，我现在这套招聘流程，我自己是真的很喜欢，至于最后成不成还得看。但我觉得这事儿挺有意思的。我的做法是这样：我基本上就在 Twitter 上发了一条，说「好，我们要招第一位工程师。你必须是那种超级超级超级拥抱 AI 的人，这是当然的；你住在哪儿我完全不在乎。我要你做的只有一件事——录一段视频，全屏、整个桌面都录下来，内容是你给一个已经存在的 app 做一个新功能。我就想看完整的全过程。而且我不想跟你面谈，不想开会，也不想先搞清楚我喜不喜欢你这个人。我只想看看你当 agent 管理者的水平如何。」收到的这些视频特别有意思，而且非常能看出我到底愿不愿意跟这个人共事。我挺喜欢这个方式的。现在已经有一批人大概会进到第二轮，第二轮我打算给他们开 Devin 的权限，让他们真刀真枪做点东西出来，同样录成视频。最后我大概会来个简单的收尾——「行，那我们该见一面聊聊」，然后把人招进来，之后我就只看他们 agent 的产出，我就这么管人。哦对，还有一点：如果他们用 Devin，你本来就能拿到 Devin 那台电脑跑起来的回放视频。所以等于又给你省了一步。
[41:37] Claire
Thank you.
谢谢。
[41:38] Claire
The one meta observation I have on this
我对这件事有一个更上层的观察，
[41:40] Claire
is you're almost hiring a human like you would hire an agent, which is like,
就是你几乎是在用招 agent 的方式招人，那感觉就像是，
[41:47] Claire
I love you. Like, we'll get along, I'm sure, but like really, I just want to see
「我很喜欢你，我们肯定处得来，但说真的，我只想看看
[41:53] Claire
how you do this. You're you're doing a verification loop is what you're saying and think about how much better this is in this insane way we used to hire.
你到底是怎么干这件事的。」你其实就是在跑一个验证循环嘛。你想想，这比我们过去那种离谱的招聘方式要好多少。
[42:00] Ryan
Yeah.
是啊。
[42:00] Claire
You like I like this person like you
以前就是那种——「我挺喜欢这个人」，然后你
[42:02] Claire
and then you're like we're going to kind of ask you to do some weird technical project and then
接着来一句「我们要给你出一道有点莫名其妙的技术题」，然后
[42:07] Claire
so very excited to see this shake out.
所以我特别期待看看这套最后能跑出什么结果。
[42:10] Claire
Okay. Well, um just to recap for everyone, the the two most AI pill on your timeline spent most of our time talking about how too much AI code doesn't really get you great product. [snorts] Amen.
好。那我给大家总结一下：你时间线上最「AI 上头」的这两个人，花了大半场时间在讲——写太多 AI 代码并不会给你带来好产品。（笑）阿门。
[42:22] Claire
You should just go outside with your human body and talk to customers.
你应该带着你的肉身走出门去，跟客户聊聊。
[42:27] Claire
Keep your priority on a piece of paper [laughter]
把你的优先级写在一张纸上（笑）
[42:30]
on your list.
写在你的清单上。
[42:33] Ryan
It works.
这招管用。
[42:33] Claire
Um, it works. But, you know, on on the technical stuff, set up your background agents. Prioritize their work in a way that you can manage. Everybody's becoming agent managers. Your coding agent can do more for you than just code. And the new interview is just let them cook and and do a screen share. Um, and also to conclude, the solo founder AI agent managers are both hiring. So [laughter] Ryan, where can we we I know we can find you online, but where can we find you and how can we be helpful? Who are you hiring for?
嗯，是管用。不过技术层面上：把你的后台 agent 搭起来；用你自己管得过来的方式给它们的活儿排优先级；每个人都在变成 agent 管理者；你的 coding agent 能替你做的远不止写代码；还有，新式面试就是放手让他们去折腾，顺便开个屏幕共享。另外做个收尾——两位单干创始人兼 AI agent 管理者现在都在招人。（笑）那 Ryan，我知道大家在网上能找到你，但具体该去哪儿找你？我们能怎么帮上忙？你在招什么岗位？
[43:07] Ryan
Thank you. So I'm super active on X, so just Ryan Carson. Um, hit me up anytime. And then uh uh if you want to help people have better divorces because people are going to get divorced, okay? Um go to untangle. us, check us out, and then just shoot me a DM with a video of you working. Um and I think we're going to hire someone pretty fast. So So move fast and uh and and say hello.
谢谢。我在 X 上非常活跃，搜 Ryan Carson 就行，随时来找我。另外——如果你想帮人把离婚这件事过得更体面一点（因为离婚这事儿总归会发生，对吧），那就上 untangle.us 看看我们，然后直接给我发一条 DM，附上一段你干活的视频。我觉得我们会很快定下人，所以动作快点，来打个招呼。
[43:32] Claire
And I'll say if you impress Ryan with your video, you're probably going to be an upcoming guest of How I AI, where you will show us how you use your background agents. Awesome. Well, Ryan, thank you so much for joining.
我再补一句：如果你的视频把 Ryan 打动了，那你八成会成为 How I AI 接下来某一期的嘉宾，来给我们演示你是怎么用后台 agent 的。太棒了。那 Ryan，非常感谢你来做客。
[43:43] Ryan
This has been a blast. Thank you.
聊得太过瘾了，谢谢。
[43:45] Claire
Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube, or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiipod.com. See you next time.
非常感谢收看。如果你喜欢这期节目，请在 YouTube 上点个赞、订阅一下；或者更好的是，留言告诉我们你的想法。这档播客你也可以在 Apple Podcasts、Spotify 或你常用的播客 App 上找到。也欢迎给我们打个分、写条评价，这能帮更多人发现这档节目。所有往期节目和节目的更多信息都在 howiipod.com。下期见。