Grok Bot's Top 14 Bots: Peng and Lauren on Agent Swarms and Trust
打开互动全文版(中英对照 + 朗读 + 问答)→Grok Bot 团队的 Peng 和 Lauren 分享各自的机器人配置——从幕僚长机器人到工程师集群,并探讨信任如何影响把工作交给智能体。
Peng and Lauren from the Grok Bot team share their personal bot setups, from a chief-of-staff bot to engineer swarms, and discuss how trust shapes delegating work to agents.
好的。大家好。今天的嘉宾是 Pen 和 Lauren,他们是 Grok Bot 团队的设计和技术人员。非常期待看到他们用于工作和个人任务的 Grok Bot 设置,以及如何使用 Grok Bot 设计和构建实际产品,还有他们看到社区制作的最令人惊讶的机器人。欢迎两位。
All right. Hey everyone. My guests today are Pen and Lauren, who are members of the design and technical staff on the Grok Bot team. Really excited to see their Grok Bot setups for work and personal tasks, and also how to use Grok Bot to design and build the actual product, and also the most surprising bots that they see the community making. Welcome both.
感谢邀请我们。
Thanks for having us.
谢谢。非常兴奋。
Thank you. Very excited.
是的,非常兴奋。我认为 Grok Bot 非常令人兴奋。我觉得这是我见过的第一个持久化的个人电脑产品,我自己每天都在用。
Yeah, very excited. I think Grok Bot is very exciting. I think it's the first persistent personal computer product that I've seen, and I use it every day myself.
但我们为什么不从这里开始呢?也许你们可以给我们展示一下你们的个人 Grok Bot 设置,以及一些你们非常自豪的机器人。也许从你开始,Pen。
But why don't we start with this? Maybe you guys can show us kind of your personal Grok Bot setup and a few of the bots that you're really proud of. Maybe let's start with you, Pen.
是的。我想分享几个我拥有的机器人。基本上,我在数字世界中几乎所有事情都在使用 Grok Bot。在高层次上,我按工作和生活分类。然后我把最常用的机器人固定在这里。Chief of Staff 是我的首选机器人。如果我不知道该由哪个机器人来处理任务,我就会把它发给 Chief of Staff。有一件事是,你几乎可以让它在数字世界中做任何事情。几天前我让它购买 3D 打印耗材,它不仅帮我买了,我还让它添加到库存清单中。所以我有一个 Notion 数据库来跟踪我所有的库存,不同颜色的耗材,一旦它们购买了,它还会为我更新列表。
Yeah. So I'd like to share a couple of bots I have. So basically I'm using Grok Bot for almost everything in the digital world I'm having. And at a high level, I categorize them by work and life. And then I have those most frequently used bots pinned here. Chief of Staff is my go-to bot. If I don't know which one should take the task, then I'll send it to Chief of Staff. And one thing is you can almost ask it to do anything in the digital world. I was asking it to buy 3D printing filament a couple days ago, and it does not only buy it for me. I also ask it to add it to the inventory list. So I have this Notion database to track all my inventory, different colors of the filament, and once they bought it, it also updates the list for me.
我刚刚发现将不同任务串联起来的优势非常强大,因为另一个例子是,假设我想在 Facebook Marketplace 上出售我的 DJI 麦克风。我只需发送一张照片,然后告诉它做几件事。所以像卖东西、在 Facebook Marketplace 上列出物品听起来很简单,但实际上有很多步骤,比如你需要查看官方网站上该型号的最新价格。我想找出我所在地区的最低报价,比如我在和谁竞争。我想了解我过去如何写商品描述,还要发布这个新列表。然后如果没有人感兴趣,每周自动降价 5 美元。所以你看,这就像我在一个提示中指定的一系列任务,它会去执行。非常有趣的是,它知道什么时候该写描述时,它会咨询我的写作机器人。它给写作机器人发消息,问它我过去如何写市场列表描述,然后它就能进行研究,找出其他人如何销售,最终列出我的列表。所以过去我们可能自己手动完成所有这五个步骤,但现在不仅可以委托单个任务给它,还可以让它为你串联任务。
I just discovered the benefit of trying different tasks together is really powerful because another example is like let's say I want to sell my DJI mic on Facebook Marketplace. I just send it a photo and then tell it to do a couple things. So like sell something, list something on Facebook Marketplace sounds easy, but actually there are a lot of steps like you need to check the latest price for the model on the official website. I wanted to find out what's the lowest price offered in my area, like who I'm competing with. I wanted to learn like how I write listing descriptions in the past and also post this new listing. And then if nobody's interested in it, automatically lower the price by $5 every week. So you see it's like a series of tasks that I dedicate in one single prompt and it will go execute the task. And it's very interesting that it knows like when it should write the description, it consults my writer bot. It messages the writer bot, it asks like how I wrote marketplace listing description in the past, and then it will be able to do the research, it'll find out like how other people are selling it, and then eventually pull up my list. So in the past we may do all those five steps manually by ourselves, but now not only can delegate individual task to it, but also ask it to train the tasks for you.
我的机器人正在取代我的电子邮件。所以我会做这样的事情:嘿,我今天收件箱里有哪些行动项?我的电子邮件仍然难以管理,对于我确定它可以为我做的事情,比如我需要付钱给朋友晚餐,我就让它去做。如果它没有访问权限,它会为我提示这个登录页面,我可以接管并在它的电脑上登录,然后它就会为我做。它正在取代我的日历。所以如果我现在和别人喝咖啡聊天,基本上任何编辑、任何事件创建、查找预订比如预订会议室,任何事情我只需放一张截图,然后让它去做。让它把事件添加到日历、修改它,诸如此类。
My bot is replacing my email. So I would do things like hey what are the action items in my inbox today? My email is still out of manage and like for things I identified that can do it for me. For example, I need to pay my friend money for dinner. I'll just ask it to do so. And if it doesn't have access, it will prompt this sign-in page for me and I can take over and sign in on its computer and it will do it for me. It's replacing my calendar. So if I'm having a coffee chat with somebody now, like basically any edits, any event creation, find booking like book conference rooms, anything I just put a screenshot and then let it do it. Let it like add the events to the calendar, modify it, stuff like that.
我注意到你有一个——好的,所以你有一个——好的,让我们看看你固定的机器人。所以你固定了一个设计师和一个写作机器人。所以设计师——所以写作机器人就像是你的写作风格,设计师就像是——是的,也许可以快速介绍一下这些。
I noticed you have a — Okay, so you have a — Okay, let's go up to your pinned bots. So you have a designer and a writer bot pinned. So designer — So the writer is like your writing style and designer is like — Yeah, maybe can quickly walk through those.
是的。所以这是一个演示账户,我这里没有所有东西,但和我自己账户中的设置完全一样。比如那三个是最常用的,比如 Chief of Staff 会管理所有其他机器人。设计师——我每天都待在设计工具里,然后设计师是我的首选机器人。我会让它和我一起头脑风暴,或者让它为我衍生出不同的设计选项。我会和它讨论,不仅和我一起工作,还执行想法。然后对于写作机器人,目前因为我不是英语母语者,很多电子邮件和 Slack 消息,我会让写作机器人润色后再发送。我们有 Slack 和电子邮件聊天小部件。所以你实际上可以看到这里的草稿,你可以审阅并发送,决定如何修改等。
Yeah. So this is a demo account that I don't have everything here, but it's the exact same setup I have in my own account. For like those three are the most frequently used like Chief of Staff going to manage all other bots. And the designer is like I stay in design tools every day and then designer is my go-to bot. I'll have it like brainstorm with me or like ask it to spin off different design options for me. And I'll just like discuss with it and not only like work with me on it, but also execute the ideas. And then for the writer bots, currently like I'm not an English native speaker. And then a lot of emails and Slack messages, I would have like the writer bot to polish it before I send it. And we have the Slack and email chat widget. So you literally see the draft that inline here that you can review and send and decide how to modify etc.
本集由 Granola 赞助。如果你连续开会,你知道实时记笔记并在会后整理有多费事。这就是我喜欢 Granola 的原因,它是市场上最好的 AI 会议笔记应用。以下是我的使用方式。Granola 在会议期间自动记笔记,我也可以添加自己的笔记。会议结束后,我使用 Granola 的配方以我想要的格式提取清晰的要点和后续步骤。然后我可以直接在 Slack 中与同事分享她的笔记,甚至让 Granola 自动分享她的笔记。老实说,在我使用的所有 AI 应用中,Granola 是最节省我时间的。现在就在 granola.ai/peter 尝试,并使用代码 Peter 注册,获得三个月免费。那是 granola.ai/peter。现在,回到我们的节目。
This episode is brought to you by Granola. If you're in back-to-back meetings, you know how much work it is to take notes live and clean them up afterwards. That's why I love Granola, the best AI meeting notes app in the market. Here's how I use it. Granola automatically takes notes during a meeting and I can add my own notes, too. After the meeting ends, I use a Granola recipe to extract clear takeaways and next steps in the exact format that I want. Then I can just share her notes directly in Slack with my colleagues or even get Granola to share her notes automatically. Honestly, of all the AI apps that I use, Granola is the one that saves me the most time. Try it now at granola.ai/peter and use the code Peter to sign up and get three months free. That's granola.ai/peter. Now, back to our episode.
你还有一个像 MeNow 应用之类的频道,里面有一堆机器人在互相交谈。
You also have a channel like me now app or something where you have a bunch of bots talking to each other.
是的。那是什么呢?一旦你有更大的任务,你希望不同角色的机器人有自己的视角并为之贡献。
Yeah. So what is that? Once you have bigger tasks that you want different bots with different roles have their own perspective and contribute to it.
这是一个例子,比如假设我们正在构建一个可以将猫的喵喵声翻译成英语的应用,我想从不同的角度开始审查这个想法。
This is an example like if let's say we are building an app that can translate cat meow to English and I want to start review this idea from different perspectives.
我会让我的 PM、设计,比如三个独立的机器人,把它们加进同一个群聊,然后我抛出这个提示词,它们就会开始讨论、围绕这件事头脑风暴。看它们怎么想这个点子、怎么互相辩论,是很有意思的。如果你想让某个特定的机器人回复你,只要 @ 它一下,就只有那个机器人会回你。如果你只是在群里发一条消息,就像你在群聊里给朋友发消息那样,它们会看一遍,然后回复——它们会自己决定要不要回。
So I will have my PM design, like, three individual bots, add them into a single group chat, and I will throw this prompt, and then they'll start to discuss and ideate about this thing. And it's interesting to see how they think about the idea, how they debate with each other. If you want a specific bot to reply to you, just simply at-mention it, and only that bot will reply to you. If you just send a message in the group chat, just like how you message your friends in a group chat, they will review it and then they will reply — they will decide if they should reply from their own end.
是啊,这真的很有意思。我觉得在别的某些 AI 应用里不太容易做到这一点,对吧?因为保持上下文干净,同时让这些各自拥有自己上下文的机器人互相聊天,是相当独特的。
Yeah, this is really interesting. I don't think this is easy to do in some of the other AI apps, right? Because keeping the context clean and having all these bots that have their own context chat with each other is pretty unique.
是的。所以你可以从两个角度来看这个问题。有些人按单个会话、任务和项目来管理自己的任务和工作。而我们是在把机器人做成更持久的、有名字的智能体,它们有自己的身份、记忆系统、自己的工具。其实就是把可复用的东西抽象成一行一行的,然后你可以为它们创建单独的聊天和群聊,让它们一起工作。比如我同一个设计师机器人有多个项目,我可以为另一个项目再开一个和它的聊天,对吧?然后我可以为我管理的不同项目反复这么做。因为我们注意到的一点是,在 AI 智能体产品里,很多时候那些聊天都非常短暂,几乎是一次性的。就像你并不在意——也许你只在意你正在积极做的那三到五个,而两周前的那些聊天,你再也不会回去看。所以我觉得这是平衡两者的好办法:凡是可复用的能力、记忆,你就把它们抽象成不同的行,而对于有截止日期的单个项目,你可以创建专门的会话,在里面和你的机器人一起工作。
Yeah. So there's like two angles you can think about the problem. Some people manage their tasks and jobs by individual sessions, tasks and projects. And we are creating bots as more persistent, named agents that have their own identity, memory system, their own tools. It's really like, abstract what's reusable into the row, and then you can have them work together by creating separate chats and group chats for them. Like even I have multiple projects with the same designer bot — I can create another chat with it for a different project, right? And then I can do this repeatedly for different projects I'm managing. Because one thing we noticed is a lot of the time in AI agent products, those chats are very ephemeral, they're almost disposable. Like you don't care — maybe you only care about the top three to five that you're actively working on, and then for the chats like you had two weeks ago, you never go back. So I think this is a great way to balance the two: for whatever is reusable, the capability, the memory, you kind of abstract them into different rows, and then for individual projects that have a deadline, you can create these dedicated sessions and work with your bots in it.
但就像一个人一样——设计师机器人,比如我跟它聊邮件应用,又聊 bark 应用,它在自己的上下文里两个都知道,对吧?是的。还是说它们是分开的?
And but just like a human — the designer bot, like if I talk to the designer bot for the mail app and the bark app, it knows about both, right, in its context? Yes. Or is it separate?
是的。所以当一个机器人在群聊里工作时,你能在它自己的私聊里看到工作状态发生了变化。所以它是同一个实体。是同一个机器人在处理你的任务。
Yes. So when a bot is working in a group chat, you will be able to see the work status has changed in its own DM. So it's the same entity. It's the same bot working on your task.
明白了。
Got it.
是的。然后我还有别的用例,我觉得挺有意思的。比如 Grok Bot 现在在替代我的播客工具,因为我听很多播客,但有时候我读到一篇文章,我想在开车或洗澡的时候听。所以我就把这个链接发给我的播客机器人,它会把它转成一个播客,我随时发一下就能听。
Yeah. And then I have other use cases I think are pretty fun. Like Grok Bot is replacing my podcast tool right now, because I listen to a lot of podcasts, but sometimes when I read an article I want to listen to it when I'm driving or taking a shower. So I'll just simply send this link to my podcast bot and it will translate it into a podcast that I can listen to anytime I just post.
所以看到它变成一种自动的、终极的不同文件格式转换器,这非常有意思。
So it's very interesting that you see it become like an auto, ultimate translator for different file formats.
另外我还想分享这个个人项目,就是我搭的那条流水线。故事是这样的:大概两周前我在旧金山唐人街等公交的时候,拍了一张一家果蔬店的照片,然后用 AI 把它变成了一个 3D 黏土微缩模型,我想“哦,这挺好玩”。我把它放到我的个人网站上——那是个非常极简的网站——我把它放在这儿,然后写了一行字“Lasting at a JBC market in Chinatown”。然后我想,哦,这可以变成一个打卡应用,因为个人网站嘛,你放上去,人们来看一次,然后就再也没人回来了——我自己都不怎么常去看。所以我就有了把它做成一个互联网打卡应用的想法,这样每次我打卡,它就能自动把位置、把那家店或商铺之类的发上去。于是我开始跟 Grok Bot 聊这个想法,最终我把它落到这条流水线上:如果我给它发一张我去过的地方的图片,还有我跟谁在一起、那个人的社交账号,它就能生成两张图,一张日间模式一张浅色模式,还能去掉背景,最后发到我的网站上,大概就是这样。
And also I like to share this personal project, like the pipeline I set up. The story was I took a photo of a produce shop in San Francisco Chinatown like two weeks ago when I was waiting for the bus, and then I turned it into like a 3D clay miniature using AI, and I thought, "Oh, this was fun." And I put it onto my personal website — it was a very minimal website — and I put it here, and then I wrote the line "Lasting at a JBC market in Chinatown." And then I thought, oh, this could become a check-in app, because for a personal website, like you put it up there and people visit it once and then nobody comes back again — like I don't even visit it that often. So I had this idea of making it an internet check-in app, so every time I check in it can just automatically post the location and post the shop or store or whatever up there. So I started to talk to Grok Bot about this idea, and eventually I was able to land it on this pipeline that if I texted it an image of the place I've been and also like who am I with, the social handle of the person, it will be able to generate two images in day and light mode and also remove the background, and then eventually post it to my website, something like this.
哦,不错。好。
Oh, nice. Okay.
它非常天马行空、好玩,就是个个人项目。但我觉得换个角度看它:现在我是把 Grok Bot 当作托管内容的主界面,而不是必须搭一个门户或 CMS 系统、在后端手动上传。很多东西都内置了——内置在我这条流水线里。这个机器人就去这个私有仓库里取最新的操作手册,然后照着步骤做,生成图片、清理干净,然后发布。
It was very whimsical, fun, like a personal project. But I think it's a different perspective to think about it: now I'm using Grok Bot as like the main interface to host things, instead of having to build a portal or CMS system and manually upload it in the backend. And a lot of it is baked in — baked in this pipeline that I have. This bot just goes and fetches the latest playbook on this private repo, and then follows the steps and then generates the image, cleans it up, and then posts it.
好。所以人们能看到你去过的所有地方。而且典型的设计师——你会考虑到日间和浅色两种模式。我都不会想到这个。是啊,这很好。
Okay. So people can see all the places you've been to. And the typical designer — you think about both day and light mode. I would not think about that. Yeah, that was good.
是的,所以这就是一个例子,你可以创造性地使用 Grok Bot,几乎把它当成你想构建的东西里的一个构建模块。而最棒的是,它能一步步引导你走向那个目标。
Yeah, so that's an example that you can use Grok Bot in a creative way, almost think of it as a building module in the thing you want to build. And the best part is it can guide you step by step towards that goal.
在我们转到 Lauren 之前——你的设计师机器人,所以你现在还在 Figma 里画框框吗,还是你让机器人做所有事,你只负责审阅?
And before we move on to Lauren — your designer bot, so are you still like drawing boxes in Figma, or you tell your bot to do everything and you just review?
我们,作为 xAI,我们用任何需要的工具来设计。我们用 Figma,用基于画布的工具,我们有自己的原型环境,有时候我们也自己提交 PR。所以设计师机器人跟我碰撞想法非常有用。我也用它做快速修改,因为它通过连接器接上了 Figma MCP,所以它能直接改画框、更新设计、使用设计 token。我有一个技能是设计师机器人能用的,所以它知道我的 Figma 文件是怎么设置的、我的设计系统是什么,以及在为我做设计时什么时候该用对的颜色、字体、间距。
We, so as xAI we do we design in whatever tools we need. We use Figma, we use canvas-based tools, we have our own prototype environments, and sometimes we also submit our own PRs. So the designer bot is really helpful to bounce ideas with me. I also make quick changes because it connects the Figma MCP through the connector, so it will be able to directly change frames and update the design and use design tokens. I have this skill that the designer bot is able to use, so it knows how my Figma file setup is, what's my design system, and when to use the right colors, typography, spacing when it creates design for me.
好。所以也许它能做前 80%,然后你可以手动调一调之类的。
Okay. So maybe it can do like the first 80% and you can manually tweak stuff or something.
是的。对。或者有时候我的工作流变成了:我做前 5%。我会把系统搭起来,做出一个关键画框,然后让它把它扩展到整个流程,做成端到端的界面。
Yeah. Yes. Or sometimes my workflow has become: I will do the first 5%. I would build the system and I'll create like one key frame, and then ask it to scale it into the whole flow and make it like end-to-end screens.
哦,太棒了。你应该像 Lauren 开源她的 pstack 那样,把你的设计技能也开源。
Oh, awesome. You should like open source your design skills like Lauren did with her pstack.
是啊。是啊。有意思。我们有这个——这些技能以我们的名字命名,比如 Lauren mode,我有个 Pen mode。
Yeah. Yeah. Interesting. Like we have this — this skills follow our names, like Lauren mode and I have a Pen mode.
Pen mode。对。正是这样。太棒了。Pen,非常感谢你分享这些。这超酷。Lauren,谢谢你等着。我也很想看看你的配置是怎样的。
Pen mode. Yeah. Think exactly. Awesome. Pen, thanks so much for sharing all this. This is super cool. Lauren, thanks for waiting. I would love to see how your setup is, too.
这可能和——
It's probably pretty different from—
很不一样。
Very different.
是的。我想分享更多工程方面的用例。我最近在做的一件事是更好地支持 Omarchy,这是一个由 DHH 和其他几位非常棒的开源人士开发的 Linux 发行版。我一直在用 Grok Bot 来控制——这是一个在 Mac 虚拟机上运行的 Omarchy 特殊版本。
Yeah. I wanted to share more of the engineering use cases. One thing I've been recently working on is adding better support for Omarchy, which is a Linux distribution that DHH and a couple of other really awesome open-source folks have been working on. I've been using Grok Bot to essentially control—this is a special build of Omarchy which runs in a virtual machine on your Mac.
它还是 alpha 阶段,但运行得相当不错。它非常强大,因为我可以让我的智能体直接编写脚本并控制虚拟机。所以,要支持这个特定的 Linux 发行版,还有什么比让 Grok Bot 直接在我的电脑上打开它并交互式安装更好的方式呢?就像它为我安装 Grok Bot 一样。你可以看到它现在正在这样做。
Which is still very alpha but it works pretty well. But it's really powerful because I can actually get my agent to literally script and control the VM. So what better way to build support for this particular flavor of Linux than to have Grok Bot actually be able to open it on my computer and interactively install it—like it installed Grok Bot for me. You can see it's actually doing that right now.
好的。所以 Grok Bot 在它自己的电脑上,但通过 Omarchy 控制你的电脑。
Okay. So Grok Bot is on its own computer but it's controlling your computer through Omarchy.
是的,所以 Grok Bot 有它自己的电脑。但它也有一个在你的电脑上运行的本地执行守护进程。所以如果你告诉它——我经常做的另一件事是:我有一个叫 Crumb 的智能体,我经常硬盘空间不够。所以我就告诉它,‘嘿,翻一遍我的硬盘,看看空间都去哪了’,然后它就去找到要删除的东西。但没错,根据你给的权限,它可以完全访问你的电脑。在这种情况下,因为我想试试 Omarchy,修复一些 bug 或者改进 Linux 上的体验,能让机器人控制虚拟机本身——我电脑上的 Linux 虚拟机——是非常强大的。因为这是我经常谈到的一点:让智能体能够验证自己的工作非常重要。所以这是一个例子,智能体可以在你的电脑上验证它的工作,当然它也有自己的电脑,你可以看看它在做什么。但没错,这是我主要的用例之一——在 Grok 上做大量的工程工作。
Yeah, so Grok Bot has its own computer. But it also has a local execution daemon that runs on your computer. So if you tell it to—another thing I often do is this: I have this agent called Crumb and I often just run out of hard disk space. So I'll just tell it, 'Hey, go through my hard disk and find where all the space is going,' and it just goes off and finds stuff to delete. But yeah, it has full access to your computer depending on the permissions you give it. In this case, because I wanted to try out Omarchy and fix some bugs or work on the experience a little bit for Linux, it was very powerful to actually have the bot control the virtual machine itself, the Linux virtual machine on my computer. Because that's one thing I always talk about: the importance of actually having the agent be able to verify its own work. So this is one example where the agent can verify its work on your computer in addition to, of course, having its own computer that you can take a look at and see what it's doing. But yeah, that's one of the main use cases I have—doing a lot of engineering work on Grok.
我管理的另一件事是我自己的个人插件栈 Pstack。所以我在这插件上做的所有工作基本都在这里进行。我会启动云端智能体,Cursor 云端智能体,它们本质上会为我做评估。所以为了对我的一套技能做出我能信任的更改——确保改进确实能带来更好的表现——我在 Pstack 里有一个评估手册,教智能体如何——它会生成几个不同模型的子智能体。它会尝试提示词。它会确保我们实际设定的意图和目标得以实现,然后只有在实现的情况下,它才会真正合并那个拉取请求。
Another thing I manage is my own personal stack of plugins, Pstack. So all of the work I do on this plugin basically happens in here. I'll kick off cloud agents, Cursor cloud agents, which will essentially do eval for me. So in order to make a change to my set of skills that I can trust—that the improvements are actually going to result in better performance—I have an eval playbook in Pstack that teaches the agent how to—it'll spawn a couple of sub-agents of different models. It'll try out the prompt. It will make sure that the intent, the goal that we actually set out to achieve was achieved, and then only if it does achieve it, it will actually land that pull request.
是的,这就是我做的大量工程工作。
Yeah, so that's quite a lot of the engineering work that I do.
好的。我注意到你的 Grok 机器人名字和 Peng 的 Grok 很不一样。它们更有性格,不只是,你知道。
Okay. And I noticed your Grok Bots have very different names from Peng's Grok. They have more personality than just, you know.
是的。我问了我的——我有一个叫 Dr. Eggbot 的机器人,我也在社交媒体上分享过。Dr. Eggbot 本质上是我设计师机器人。我的意思是,不是设计师机器人,而是机器人设计师机器人,它为我设计机器人,带着我期望从我使用的机器人身上得到的所有严谨性、工程严谨性。我还让它想食物主题的名字。我不知道。我猜这像是亚洲人的特点。我们热爱食物,所以得给我的机器人起可爱的名字,比如大福饺子。所以挺有趣的。
Yeah. I asked my—so I have this bot called Dr. Eggbot, which I've also shared on my socials. Dr. Eggbot is essentially my designer bot. I mean, not a designer bot, but a bot designer bot who designs bots for me with all of the rigor, engineering rigor that I expect from the bots that I use. And I also get it to come up with food-themed names. I don't know. I guess it's like an Asian thing. We love food, so we got to call my bots cute names like Daifuku Gyoza. So that's pretty fun.
然后和 Peng 类似,我也有一个幕僚长。我有一个个人助理。我的个人助理帮我处理午餐、航班、预订酒店之类的事情。有时我忽略她,然后她就有点俏皮。然后我的幕僚长帮我处理各种随机的事情。但如果你看我的侧边栏,我这里有一堆工程师机器人。我还有一些社交机器人来管理我的社交媒体。所以我做的很多工作都通过这两个或三个智能体,我猜。要么我在和幕僚长说话,我的幕僚长和我这里的负责人 Matcha 合作,要么如果是 Suki,我的个人助理,她可能会为我订航班。我还没有建立个人助理团队,因为我没有那么多——我猜我没那么忙,不需要一队机器人来帮我。到目前为止她做得非常好。
And then similar to Peng, I also have a chief of staff. I have a personal assistant. My personal assistant helps me with things like lunch, my flights, booking hotels. Sometimes I ignore her and then she gets a little bit sassy. Then my chief of staff helps me with just all sorts of random things. But if you see in my sidebar, I have a bunch of engineer bots here. I have some social bots as well that manage my social media. So a lot of the work that I do goes through these two or these three agents, I guess. Like either I'm talking to the chief of staff and my chief of staff is working with my in-lead Matcha here, or if it's Suki, my personal assistant, she might book flights for me. I haven't set up a personal assistant team yet because I don't have that many—I'm not that busy, I guess, that I need a team of bots helping me there. So far she's been doing a really great job.
然后 Dr. Eggbot 会在我开始意识到‘哦,我开始用一个智能体做很多事情’时为我设计机器人,这有点——我会说这意味着我给那个智能体添加了很多非常多样和不同的上下文,这可能会让人困惑。所以 Dr. Eggbot 做的一件事是它有一个例程——这里,这是 Dr. Eggbot 会为你设置的一个例程,它会查看你所有的机器人以及你和它们聊天的记录,然后提出建议,比如‘这个机器人老是做这个不太好的事情’,然后它会建议一些改进,无论是技能还是新机器人还是调整例程。它还会检查你现有的例程,确保它们高效,因为例程实际上相当——如果你让它们运行得太频繁,对你的使用量来说可能相当昂贵,因为它会不断唤醒智能体。比如如果你每 10 分钟一次,那就是很多次唤醒。所以 Dr. Eggbot 在这方面帮助很大。
And then Dr. Eggbot just comes up with bots for me whenever I'm starting to realize, 'Oh, I'm starting to use one agent for a lot of things,' and that's kind of—I'll say this means I'm adding a lot of context to that agent that's very diverse and very different, which can be maybe confusing. So one of the things that Dr. Eggbot does as well is it has a routine where—here, this is a routine that Dr. Eggbot will set up for you where it'll look through all of the bots that you have and the transcripts that you've been chatting with them about, and it'll propose things like, 'This bot keeps doing this thing that's not so good,' and then it'll suggest some improvements, whether it's skills or a new bot or adjustments to routines. It also goes through your existing routines as well and makes sure that they're efficient because routines are actually quite—if you have them run too often it can be quite expensive on your usage because it just keeps waking up the agent. Like if you do every 10 minutes, that's a lot of wakeups. So Dr. Eggbot does help a lot with that.
好的。那么,好吧,我要给你几个问题。所以对于 Suki 机器人——
Okay. Well, okay, I'm going to give you a couple questions. So for Suki bot—
当然。
Sure.
所以你实际上依赖它来订航班之类的吗?你真的依赖它做大额购买吗?你给它最大的自主权是什么?
So what do you actually rely on it to book flights and stuff like that? Do you actually rely on it to make big purchases? What's the most autonomy you've given it?
是的,我实际上订了我的——我很快要在伦敦和阿姆斯特丹做一个演讲,是 Compile on the Road 会议的一部分。希望我不会泄露我的航班信息,但基本上我通过 Tuki 预订了整个行程。这实际上相当——我对她的做法印象相当深刻,因为这是一个非常复杂的多段航班。我不知道你们怎么称呼它。这不是一个简单的从橙县到伦敦的往返航班。而是橙县到丹佛到伦敦到阿姆斯特丹,然后阿姆斯特丹返回我家。
So yeah, I've actually booked my—so I'm going to give a talk soon in London and Amsterdam as part of the Compile on the Road conference. And so hopefully I don't leak my flight details, but basically I booked the entire trip through Tuki. And this was actually quite—I was actually quite impressed by how she did it because it was a very complex multi-step flight. I don't know what you call it. It's not a simple return flight from Orange County to London. It was like Orange County to Denver to London to Amsterdam and then Amsterdam back to my home.
是的。
Yeah.
我完全通过——我们使用这个,叫什么来着,Navan,我想是。
And I booked it entirely through—we use this, what's it called, Navan, I guess.
好的。公司用的东西。是的。
Okay. Corporate thing. Yeah.
是的。这就像是一个公司用的订票工具。但它完成了所有工作,找到了最佳的航段和我能获得的最佳舱位、最佳时间,还为我找到了所有酒店,都在活动附近。所以那真是——它完全自主地完成了。我只是给了它一个 Slack 消息的链接,里面有一些会议的细节,然后说帮我订一些酒店和航班,并在你点击批准之前和我确认——抱歉,是在你点击预订之前——然后它给出了一些建议。我就点了是,它就为我预订了一切。所以那非常非常好。这可能是我通过 GOP 做的最大一笔购买,我想。
Yeah. It's like a corporate thing to book flights for you. But it did all of the work in finding the best legs and the best class of flight that I could get away with, the best timing, and it also figured out all the hotels for me, which were in close proximity to the event. So that was really—it did it all entirely autonomously. I just gave it a link to a Slack message with some of the details about the conference and said book some hotels and flights for me and check in with me before you hit approve—sorry, if you hit book—and then it came up with some proposals. And I just hit yes and it just booked everything for me. So that was very, very nice. It's probably the biggest purchase I've done through GOP, I guess.
好的。但你仍然需要向聊天指出它计划预订什么之类的,对吧?在你——
Okay. But you still kind of got to like point out to chat what it plans to book and stuff, right? Before you—
是的,它还没有——是的,它还没有到那个程度——我想我不想要疯狂的红眼航班或者被安排在飞机后部。所以我想先批准。
Yeah, it wasn't—yeah, it wasn't to the point yet where—I guess I didn't want to have some crazy redeye flight or be put in the back of the plane. So I wanted to approve it first.
是的,我有一个非常类似的机器人——我正在尝试再次预订去日本的假期。你可以在 Google Flights 上设置价格提醒,但机器人更聪明,因为它能找到我没有设置的其他路线之类的。所以它最近实际上帮我省了一大笔钱。
Yeah, I have a very similar bot that—I'm trying to book a vacation to Japan again. And you can set up price alerts on Google Flights, but the bot's smarter because it can find other routes and stuff that I haven't set up. So it actually saved me a bunch of money recently.
是的。
Yeah.
是的,它绝对可以。是的,我也用它来买我的个人物品。比如我们卫生纸用完了。我就告诉它,嘿,重新下单我们的亚马逊卫生纸订单,它就去做了。
Yeah, it definitely can. Yeah, I also use it to buy my personal things. Like we're running out of toilet paper, for example. I'll just tell it, hey, reorder our Amazon order for toilet paper and it just goes off and does that.
实际上有一个时刻,在 GrabBot 开发的早期,我们还在迭代它的时候,出现的一个话题是你的机器人能够给你发短信。所以我不久前给 Tuki 添加的一件事是——我基本上通过一个第三方服务给了她自己的电话号码。然后她就可以真的给我发短信,因为我非常不擅长记得吃午饭。所以我需要提醒。然后我经常会有——我有时会在我的 Mac 上开启勿扰模式。我看不到提示,但然后我收到短信,它发出声音,我就想哦对,我得订午餐了。所以我也教了 Tuki 如何使用 DoorDash,当然,给我买午餐。这样我就不用太担心了。
There was a point actually, very early in the GrabBot development where we were still iterating on it, that one of the topics that came up was your bot being able to text you. So one of the things I actually added to Tuki a while back was this—I basically gave her her own phone number through a third-party service. And then she could actually text me because one thing I'm really bad about is remembering to eat lunch. So I need the reminder. And then I often will just get—I'll just sometimes have do not disturb mode on on my Mac. And I don't see the ping, but then I get the text message and it makes a sound and like oh yeah I have to order lunch. And so I also taught Tuki how to use DoorDash, of course, and get me lunch. So I don't have to worry about it so much.
那很有趣。午餐是我一天中的亮点,所以我总是期待它——从不会忘记我的午餐。
That's funny. Lunch is the highlight of my day, so I always look forward to it—never forget my lunch.
是的。是的。有些——我不知道。你就是太专注了,你知道,然后就忘了吃。
Yeah. Yeah. Some—I don't know. You just get too in the zone, you know, and then you just forget to eat.
好的。所以我的第二个问题是:我注意到你有一个领导和两个工程师机器人,对吧?所以基本上你——
Okay. So my second question is: I noticed you have a lead and two engineer bots, right? So basically you—
我有三个。每个负责一个任务还是什么,或者你如何分配责任?
I have three. Each one works on a ticket or something, or how did you dedicate responsibility?
嗯是的,所以我实际上主要和我的幕僚长交谈,然后我的幕僚长和工程领导交谈。但我用 Dr. Eggbot 设置了这些机器人——这些工程师机器人,当然。但我推动的一件事——我告诉 Dr. Eggbot 要做的是确保工程领导的工作真的是将任务分解成更小的部分,并委托和监督其他机器人,而不是自己工作。所以你可以在描述中看到,基本上它有——甚至机器人 ID 也在里面——但它告诉它永远不要自己工作,总是委托给那四个机器人。然后它基本上就这样做。很多工作也是通过云智能体完成的。例如,其中一个机器人为了完成工作,会创建一个云智能体,然后你可以在 Cursor 或网页上打开它。
Um yeah, so I actually mostly talk to my chief of staff and then my chief of staff talks to the engineer lead. But I set these bots—these engineer bots—up with Dr. Eggbot, of course. But one of the things I kind of drove—I told Dr. Eggbot to do is to make sure that the engineer lead's job is really to break down tasks into smaller pieces and delegate and supervise other bots rather than do work on its own. So you can see in the description that basically it has—even the bot ID is in there—but it tells it to never do work on your own and always delegate to those four bots. And then it basically does that. And a lot of the work as well is done through cloud agents. So for example, one of these bots will, in order to do the work, they will create a cloud agent and then you can open that in Cursor or on the web.
好的。
Okay.
但是的,这让我基本上可以根据任务类型使用不同的模型。因为有些任务更直接,我只想要非常快的速度。其他任务可能需要更多的规划和更多的推理,然后我可能会调整推理水平。所以 GrabBot 中的云智能体给了我很多非常细粒度的控制,我想。也确保我的机器人保持——我有点把每个工程师当作他们的主管。我们有领导管理他们,但他们自己尽量不自己工作。所以他们总是为我生成云智能体。所以他们做所有的消息传递,但让我编排真正大规模的智能体集群来做——比如如果我需要处理一个非常大的项目,我可以直接和我的幕僚长说,把这个分解,我有一个巨大的项目,也许我们先在 Notion 文档中做一些规划,然后它会开始思考它需要做的工作的阶段,然后它会委托给工程领导,然后工程领导会进一步分解,让机器人处理。
But yeah, this allows me to essentially make use of different models depending on the type of task. Because some tasks are more straightforward and I just want really fast speed. Other tasks maybe require a bit more planning and a bit more reasoning, and then I might adjust the reasoning level. So cloud agents in GrabBot give me a lot of very granular control, I guess. And also ensures that my bot kind of stays—I kind of treat each of these engineers as more like their supervisors in a way. We have the lead that manages them, but then they themselves try not to do work on their own. So they always spawn cloud agents for me. So they do all the message passing but let me orchestrate really massive agent swarms to do—like if I need to tackle a very big project, I can just talk to my chief of staff and say break this down, I have a giant project, maybe let's do some planning first in a Notion doc, and then it will start thinking about the phases of the work that it needs to do and then it'll delegate it all to the engineer lead who will then break it down even further and have the bots work on that.
好的。然后每个都会带回一个 PR 之类的,机器人会先审查它。是这样的吗?
Okay. Then each one will come back with a PR and stuff, and the bots will review it first. Is that kind of how it works?
是的,基本上。不过对我来说,我经常做并且也在 X 上经常谈论的一件事是,例如,GrabBot 代码库,我们试图让它非常——怎么说——让它非常智能体友好。架构和代码库非常智能体友好。
Yeah, essentially. Although for me, one of the things that I've done a lot that I talk about a lot as well on X is the idea that, for example, the GrabBot codebase, we've tried to make it very—how do I say—make it very agent-friendly. The architecture and the codebase is very agent-friendly.
所以,很多时候它实际上只是让我自动合并我的 PR。所以,呃,说起来有点吓人,但有时我甚至直到 PR 落地后才看它,然后我就想,“哦,好吧。是的,看起来不错。”
So, a lot of times it actually just lets me automerge my PRs. So, uh sounds kind of scary to say, but sometimes I actually don't even look at the PR until after it's landed and then I'm like, "Oh, okay. Yeah, that looks good."
明白了。
Got it.
是的。
Yeah.
哇。好的。所以,所以就像你有一个软件工厂。那是那个流行词吗?
Wow. Okay. So, so like you have like a software factory. Is that is that the buzz buzz word?
呃,我喜欢叫它米其林厨房,因为我觉得软件工厂。是的。呃,是的。我喜欢叫它米其林厨房,因为对我来说,米其林厨房——你看,就像这个,呃,当你想到米其林星级厨房时,对吧,你会想到质量,对吧,以及制作那些,呃,你知道的,产品,对吧,也就是一道菜或一餐,所投入的工艺,呃,而且仍然有规模,对吧,就像你知道的,一个厨房,一个餐厅可能有,你知道的,50 个座位,或者 100 个座位,呃,所以你需要能够像大规模生产质量。所以这就是我有的理想愿景,就像我想建立一个米其林厨房,而不仅仅是一个,你知道的,工厂,呃,有趣的是,我认为很多人的,就像当你说工厂时,他们有这种内涵,就像哦,你知道的,它只是像它被大规模生产,它就像是的,粗制滥造,就像低质量
Uh, I like to call it the Michelin Kitchen because I think software factory. Yeah. Uh, yeah. I like to call it the Michelin kitchen because to me Michelin kitchen you see like there's this uh when you when you think about a Michelin starred kitchen right you think about the quality right and the craft that goes into making those uh you know the the product right which is a a dish or a meal uh and there's still scale involved right like you know a kitchen a restaurant might have you know 50 seats or or or 100 seats uh so you need to be able to like produce quality at scale. So that's the aspirational vision I have is like I want to set a Michelin kitchen not just like you know a factory which uh interestingly I think a lot of people's like when you say factory they have this connotation like oh you know it's just like it's massured it's like yeah slob is like low quality
但我一直试图追求两者兼顾,对吧,就像我们如何在大规模下获得质量,所以是的,这是,呃,我在 pstack 上做的很多工作,以及像将 Grok Bot 与它一起使用,都是沿着这条线,你知道的,我们如何创建那个米其林厨房?
but I I always try to aspire to do both right like how do we get quality at scale so yeah this is uh a lot of the work I've done on pstack and like using Grok Bot together with it is in that line of you know how do we create that Michelin kitchen?
是的,你应该弄一个,你应该做一个像 concurge 酒吧之类的东西,说话像你知道的,嘿,我们策划了这个来自当地农场的 PR,然后呃,说话像,说话像复古的
Yeah, you should get that you should make like a concurge bar or something that talks like you know, hey, we curated this PR from the local farm and uh talk like a talk like a vintage
是的。是的。是的。我们需要像手工的。这里有一个手工 PR 供你审查。
Yeah. Yeah. Yeah. We need like artisal. Here's an artisal PR for you to to review.
所以,所以基本上就像 Pun 的工作只是把你开源技能,然后以某种方式构建到产品中。这有点像 PP 试图做的吗,或者不,我只是开玩笑。就像是的。是的。不,很好。你有,你肯定有一个系统设置用于
So, so basically like Pun's job is just to like take your open source skills and somehow build into the product. Is that kind of what PP is trying to do or No, I'm just kidding. It's like Yeah. Yeah. No, it's great. You have you definitely have a to system set up for
是的。我认为对我来说有趣的事情也是,我发现有了 pstack 以及一些,所以内部我们,呃,我们让每个人,不是每个人,但很多人有自己的模式技能。所以像poteto-mode,呃,就是开源中叫的,但内部叫劳伦模式,你知道的,像 Ping 分享的有 Ping 模式,以及很多不同的,你知道的,工程师和,你知道的,设计师、产品经理都有自己的模式技能,呃,我认为这让我们真正能够复合我们已经设置在代码库中的一些技能。所以是的,一个很大的动机,呃,重新设计和架构 Grok Bot 代码库是为了让每个人,对吧,都能在高层次上贡献,而不仅仅是工程师。所以在这方面投入了很多心思。嗯,当然,是的,就像人们在 Grok Bot 中启动这些智能体,呃,有时也在 cursor 中。呃,cursor 和 Grok Bot 插件的好处是它们可以互操作。所以你可以安装 pstack,例如在 Grok Bot 和 cursor 上,并重用相同的技能。
Yeah. I think the interesting thing for me as well is like I found that with pstack and with like some of the so internally we uh we have like everyone has like not everyone but a lot of people have their own mode skills. So like poteto-mode uh is is is what is called uh in open source but then internally it's called lauren mode and you know like ping was sharing there's ping mode and like a lot of different you know engineers and you know designers PMs have their own mode skills um and I think this allowed us to really compound on some of the skills that we've already set up in the codebase. So yeah big a big motivation for uh uh of of the you know redesigning and architecting the Grok Bot codebase was so that everyone right could contribute at a high level not just engineers. So a lot of care has been taken into doing that. Um and of course yeah like people are kicking kicking these agents off in Grok Bot um as well as sometimes in cursor. Uh and the nice thing about cursor and Grok Bot plugins is that they interoperate. So you can install pstack for example on both uh grabbot as well as cursor and kind of reuse the the same skills.
有趣。是的,我没有想到这个。所以,所以你的机器人也可以调用不同的技能,对吧?就像假设我有一个像正确的像我的规模或某种规模。你只需触发它,用斜杠。
Interesting. Yeah, I didn't think about this. So So your bots can also call different skills, right? Like let's say I have like a right like me scale or some sort of scale. You just trigger it the slash.
你可以,是的,如果你做像斜杠斜杠poteto-mode,你会看到,嗯,它是,呃,你可以在那里使用那个技能。嗯,你也可以随时查看市场。嗯,我实际上不知道是否有其他插件只是技能,但 pstack 是其中之一。呃,但我们有很多插件也带有技能。所以,呃,除了像,你知道的,一个 MCP 服务器。所以,例如,就像如果你,我不知道你是否安装了 X,呃,X 插件,例如,它除了工具之外,我认为还包括一个技能。嗯,是的,就像所以
You can Yeah, if you do like slash slash poteto-mode, you'll see that um it's uh you can use that skill there. Um, and you can always look at the marketplace as well. Um, I actually don't know if there are other plugins that are just skills, but pstack is one of them. Uh, but a lot of the plugins we have also come with skills. So, uh, in addition to like, you know, an MCP server. So, for example, like if you I don't know if you install the the X uh the X plugin for example, it includes I think one skill in addition to the tool. Um, and yeah, like so
插件可以包含很多不同的功能,扩展你的 Grok Bot 实例,嗯,做很多非常有趣的事情。
plugins can contain like lots of different pieces of functionality that extend your grabbot instance um to do like lots of really interesting things.
太棒了。呃,在我们结束你的分享之前,像,嗯,我很好奇你的社交机器人,像,嗯
Awesome. Uh, before we wrap up your share, like um I'm I'm curious about your social bots like um
是的,
yeah,
你如何将信号与垃圾分开在 X 上?
how how do you separate the signal from the slop on X?
嗯,是的。所以,我的很多社交机器人连接到,呃,X 以及像 LinkedIn,呃,我想我倾向于使用 X 多于 LinkedIn。嗯,但我有,呃,我有一个机器人叫 poteto,当然,它会遍历我所有的提及。所以像,呃,我在 X 上有这个梗,如果你说 poteto 三次,我就会收到通知。呃,那就是这个机器人基本上。这个机器人就像有一个例程,它检查 X,呃,我想我设置为每 30 分钟左右,呃,检查所有提及,呃,然后我就跳进去,我经常回复,呃,或者它只是聚合,就像你知道的,比如五个用户报告了同样的 bug,对吧,也许你应该看看那个,所以我有点把社交网络看作不仅仅是管理我的,我想我的社交网络,但也作为我的外循环的一种形式,因为我也在社交网络上做一些工作。呃,这意味着你知道的,我可以保持在脉搏上,你知道的,人们在报告 bug 或遇到问题,或者你知道的,也许我们发布的新功能令人困惑,你知道的,像我希望能够
Well, yeah. So, a lot of my social bots are connected to uh X as well as like LinkedIn uh which I guess I I tend to use X more than LinkedIn. Um but I have u uh I have like one bot called Potato of course who goes through all of the all of my mentions. So like uh I have this like meme on X where if you say poteto three times I get like a notification. Uh and that that that's this bot basically. this bot is like has a routine where it's checking x uh I think I set it to every 30 minutes or so uh to check for all the mentions uh and then I just jump in there and I'll often respond uh or it'll just kind of like aggregate like you know like you know five users have reported the same bug right and maybe you should take a look at that and so I I kind of think of the socials as not just not just managing my I guess my social uh networks but also as a form of like my outer loop where since I'm using the socials a bit a bit for work as well. Uh it means that you know I can stay on top of like the pulse of you know people are reporting bugs or they're running into issues or you know maybe uh a new feature that we launched is confusing you know like I want to be able to
呃,像保持连接,并让那反馈给我的内循环,你知道的,我的工程师,因为这些都是机器人,对吧,我可以,我总是可以说,像我可以去像 matcha 说,像嘿,呃,你知道的,呃,poteto 说了 blah blah blah。呃,然后上下文就可以在那里抓取,对吧?我不必重复我得到的所有 X 上下文。我不必复制粘贴任何东西。我只说去和土豆谈谈,弄清楚,对吧?然后带着如何修复这个问题的设计提案回来给我。嗯。
uh like stay connected to that and have that feed my inner loop of you know my engineers cuz then because these are all just bots right I can I can always say like I can go to like matcha and say like like hey uh you know uh poteto said blah blah blah Uh, and then the context can just be grabbed there, right? I don't have to repeat all of the X context that I got. I don't have to copy paste anything. I just say go talk to poteto and figure it out, right? And come back to me with a design proposal for how we fix this issue. Mh.
所以我认为,是的,当你,对我来说,grapa 的力量在于当你将所有这些不同的事物连接在一起,让它们有点,呃,我不知道是否有更好的词,但协同,你知道的,上下文可以从很多不同的地方抓取,嗯,并以一种比我全部自己打出来要丰富得多的方式组合在一起。
So I think like yeah when you for me the the the the power of grapa comes when you connect all of these different things together and have them kind of uh I don't know if there's a better word for like but synergize you know like the the context can can be grabbed from lots of different places um and kind of be grouped together in a way that is much richer than if I were to type it all out by myself.
是的。是的。我的意思是,上下文是这一切 AI 东西的一切,嗯,是的。
Yeah. Yeah. I mean like context is everything with this AI stuff and um Yeah.
是的。基本上所有这些机器人都住在同一个房子里,它们可以互相交谈,像你知道的,互相传递意识,对吧?
Yeah. Basically all all these bots live in the same house and they can talk to each other and like you know pass conscious to each other, right?
是的。
Yeah.
酷。呃,还有,呃,土豆来自哪里?你只是真的很喜欢土豆吗?
Cool. Uh and uh where where does potato come from? Do you just really like potatoes?
呃,是的,它就像是一个,我甚至不知道。
Uh yeah, it was just like a I don't even know.
我想那是很多年前了,我当时在玩一个 MMO,想不出名字。我对食物有点执念,你从我所有机器人的名字就能看出来。所以我就想,哦,土豆,为什么不呢?我当时也没怎么认真想。结果正常拼写的 potato 已经被占了。
I think it was many years ago, and I was just playing an MMO, and I couldn't come up with a name. I have a thing for food, as you can tell from all my bot names. So I was like, oh, potato, why not? I wasn't really thinking that hard. And then potato with the normal spelling was taken.
嗯。
Mhm.
所以我就想,哦,好吧。我当时其实在学日语。所以我就想,哦,好吧,也许你可以用日语片假名来拼 potato。从那以后就这么一直用下来了。
So I was like, oh, okay. I was actually learning Japanese at the time. So I was like, oh, okay, maybe you spell it like potato in Japanese katakana. And that kind of stuck ever since.
明白了,明白了。好。行,这非常有帮助。我只想问几个收尾问题。我看着 Grok Bot 的界面,感觉它几乎像是 Slack 的一个版本,你有私信、有队友,对吧?然后你可以把他们拉进群聊之类的。你不需要分享任何秘密,但网上有很多讨论说这东西还是很“单人模式”的 AI,就像一个人对着一堆机器人说话。也许你可以聊聊怎么让这东西变成你、Lauren 和 Peng 也能一起对着一堆机器人说话,或者你们有兴趣做这个吗?还是 Peng 你想聊聊?
Got it. Got it. Okay. Cool. All right. Well, this is super helpful. And I just want to ask a few wrap-up questions. So I feel like when I look at the Grok Bot interface, it's almost like a version of Slack where you have your DMs, your teammates, right? And then you can add them to your group chat and stuff. And you don't have to share any secrets, but there's been a lot of stuff online about how the stuff is still very single-player AI. It's like one human talking to a bunch of bots. Maybe you can talk a little bit about how you can make the stuff so that you and Lauren and Peng can also talk to a bunch of bots together, or is there any interest in doing that? Or Peng, you want to talk about it?
是的,我觉得我们开始设计 Grok Bot 时,出发点是:怎么让人们从手动管理任务中毕业?因为在我看来工作分三个层级。过去我们自己在某个环境里干活——在 Figma 里画矩形,在 Ocean 里打字,在 Slack 里发消息。现在人们已经进入了一个阶段,我们在提示 AI 替我们做这些。但在你看到的很多当前设置里,我相信我们正走向未来:人们不再手动管理所有这些机器人,而是开始设计系统、设计护栏,怎么让机器人与其他人、与你一起自主协作。这就是那种能动性和智能被注入进去。
Yeah, I think when we started designing Grok Bot, we started from a mindset of: how can we enable people to graduate from manually managing the task themselves? Because personally I see it as three different tiers of working. In the past, we work in a context ourselves — we draw rectangles in Figma, we type text in Ocean, we send messages in Slack. And now people already got into a stage where we are prompting AI to do that for us. But in a lot of setups that you've seen people doing right now, I believe we are moving towards the future where, instead of manually managing all those bots, people start designing the system, designing the guardrails, how do you make the bots work together with other humans, with you, autonomously. That was that agency and intelligence ingested.
很多 AI 智能体产品非常以模型为中心,你从 UI 就能看出来它很模型导向。我们在模型外面搭了一层 harness,然后在 harness 之上做产品,所以界面是为那量身定制的。在这个过程中我们造了很多技术黑话和概念,我妈永远不会知道什么是 solid MD。所以这是一个实验,看我们怎么让概念、系统模型去匹配用户的心智模型。然后我们发现这种拟人化、有名字的智能体效果很好,能引起共鸣,因为这就是你今天在物理世界里运作的方式。你不需要知道车险公司怎么运作,但你知道该打给谁——你有一个代理人。如果你出了车祸,你不需要知道医疗系统怎么运作,但只要你有了医保,你就有那个医生诊所,现实世界里所有这些基础设施。所以让那个接触点成为你主要的输入输出界面,对我来说是一种范式转变。
A lot of AI agent products were very model-centric, because you can see from the UI it's very model-oriented. We built a harness around the model, and then we build product around those harnesses, so the interface was tailored for that. Along the way we created a lot of technical jargon and concepts that my mom would never know — what is a solid MD? So it's an experiment to see how we can make the concept, the system model, match the user's mental model. And then we found this personified, named agents worked pretty well, and it resonates with people because that's how you operate with the physical world today. You don't need to know how the car insurance company works, but you know who to call — you have an agent for that. If you have a car accident, you don't need to know how the medical system is working, but as soon as you have health insurance, you have this doctor's office, all those infrastructure in the real world. So having that point of contact become your main input-output interface is a paradigm shift for me.
我们看到内部工作流变成了:我抓到一堆 bug,然后我就截图、标注,丢给一个机器人,它会创建并发布到一个 Notion 页面,我工程师的机器人会从那里接手。所以有点像我们开始以一种非常间接、智能体式的方式使用这个应用。但不知怎么的,它能把所有点连起来,然后以一种非常自主、长时间运行的方式有机地长成新的工作流。
We see an internal workflow become: I catch a bunch of bugs, then I just take screenshots and annotate them, and then dump them into a bot, and it will create and post it into this Notion page that my engineers' bots would pick it up from there. So it's kind of like we started to use the app in a very indirect or agentic way. But somehow it was able to connect all the dots together and then organically grow into new workflows in a very autonomous and long-running way.
也许这是使用这东西的通用最佳实践,因为某种程度上人类还在回路里——我们也在把上下文从一个传给另一个,试着提供反馈。我不知道,我总是有点怀疑当人们说,你知道,我就去睡觉,让这个机器人做所有事,就做个数据设计或者给我建个应用,我回来它就完美了。但我想你们可能更有经验。你们有没有什么技巧,让这些机器人跑得更久而不跑偏,或者真的产出你想要的东西,除了用 Lauren 的技能之外?还有别的技巧吗?
Maybe that's a general best practice overall for using this stuff, because in some ways the human is in the loop — we're also passing context from one to another and trying to provide feedback. And I don't know, I'm always a little bit skeptical when people say, you know, I'm just going to go to bed and let this bot do everything, just make a data design or build me an app, and I'll come back and it's perfect. But I guess maybe you guys have more experience doing this. Do you guys have any tips for letting these bots run longer without going off the rails, or actually producing stuff that you want, other than using Lauren's skills? Are there other tips to do this?
有。我觉得对我来说通用的一点是——你不一定要用 pstack 之类的——但我觉得这和技能有关。我聊过很多这个,但我认为它最终回到信任,对吧?就是你愿意给你的智能体或机器人多少信任,让它替你完成一个任务?
Yeah. I think the general thing for me is — you don't have to use pstack or anything — but I think there is something to do with skills. And I talk about this a bunch, but I think it ultimately comes back to trust, right? Like how much trust are you willing to give your agent or your bot to be able to do a task for you?
是的。
Yeah.
我觉得我一直试图鼓励人们做的一件事,不管你是不是 pstack的用户,就是真的花时间去观察机器人或智能体干活,然后你纠正它。但在对话结束时,你把它变成一个技能,对吧?你把它变成一个可复用的技能。比如,也许你有一个技能,你让你的机器人填一张报销单,对吧?基于翻遍我这个月的所有开销并归档——全放进报销工具里,从我邮箱里找收据,等等等等。一旦你在那个第一轮循环上迭代了很多次,对吧,如果你想让你的技能好到可以直接打斜杠报销单,然后它就干得很好。
And I think one thing I always try to encourage people to do, whether you are a pstack user or not, is to actually spend the time to observe the bot or agent doing the work, and then you course-correct it. But then at the end of the conversation, you turn that into a skill, right? You turn that into a reusable skill. So for example, maybe you have a skill where you ask your bot to fill up an expense report form, right? Based on going through all of my expenses this month and filing the — put it all into the expense tool, find the receipts from my email, blah blah blah. And once you've kind of iterated a lot on that first loop, right, if you want to be able to get your skill good enough to the point where you can do like slash expense report, right, and it just does it well.
是的。
Yeah.
当你到了可以一次性搞定的地步,对吧,你只说斜杠报销单,它就替你把事办了,那其实给了你更多信心走开,说,好,现在你可以自己去自动建立一个例程,对吧?比如每次我收到一封带收据的新邮件,就自动帮我归档这份报销单。你就不用那么盯着它了。
And when you've gotten to a point where you can just do a one-shot, right, you just say slash expense report and it just does the thing for you, then that actually gives you much more confidence to walk away and say, okay, now you can just go off and automatically set up a routine, right? Like every time I have a new email come in that has a receipt, automatically file this report for me. And you don't have to babysit it as much.
但我觉得这没什么秘诀。在我看来,无论是做设计工作、工程工作还是生活管理,归根结底都取决于你对你的机器人有多信任。如果你还不信任它,那么像技能、提示词这些真的能帮你建立信任。但没错,我觉得这必须循序渐进。很难让一个可能从未用过机器人或没有设置很多技能的人说:‘好了,现在你可以放手了。’当然,他们无法信任它,对吧?因为他们还没见过它能做什么。但一旦你和智能体一起做过一两次,并且知道它们可以重复,那么你就可以真正解放自己,去做其他事情,对吧?因为现在你可以信任你的机器人会做正确的事。
But I think there's no secret to that. In my opinion, whether it's doing design work or engineering work or life admin work, it all comes back to how much trust you have in your bot. And if you don't trust it yet, then there are things like skills, prompting, that kind of really help you build that trust. But yeah, I think it has to be gradual. It's very hard to get someone who's maybe never used a bot before or they don't have a lot of skills set up and then say, 'Okay, now just, you know, you can walk away.' Of course, they're not going to be able to trust it, right? Because they haven't seen it do what it can do. But once you've done it once or twice with the agent and you know that they can repeat that, then you can actually free yourself up and go do other things, right? Because now you can trust your bot will do the right thing.
所以,好吧。那么,我猜最后一个问题是,如果我刚下载 Grok Bot,那么步骤大概是:为你的手动工作流设置一堆机器人,手动过一遍,然后如果你能进入舒适状态,就告诉它们扩大规模,最后设置一些例程来主动执行。这大概是你们推荐的吗?
So, okay. So, okay. Okay, so I guess last question is if I just download Grok Bot, then it sounds like the steps are maybe set up a bunch of bots for your manual workflows, go through them manually, and then if you can do a comfortable state, tell them to build a scale and then eventually set up some routines to like do it proactively. Is that kind of what you guys recommend?
这是我的流程。我有点好奇你自己的。
That's my flow. I'm kind of curious about your own.
是的,我完全同意。这是一个非常渐进的过程,因为通常我会建议从简单开始:只给它一个任务。直到今天,我还在测试它的极限。基本上,我尝试一切——我触摸键盘和鼠标,试着把它专门交给它,看看它如何工作。所以你几乎必须过度拉伸它,看看极限在哪里。而且我会把这描述为建立你和机器人之间的关系,因为这是非常有意设计的:所有能力都是共享的——工具、连接器、技能——但记忆和偏好是每个机器人本地的。所以这有点像把那些可重用的组件抽象到角色中,帮助人们管理,然后形成这种关系。这需要一点时间,但试试看:从简单任务开始,然后开始一个重复性任务,然后设置例程,让它自动化,做吧,也许它消耗太多 token,就调低一点,然后连接更多工具、技能,定制它,所以它逐渐变得更成熟,它了解你。所以这甚至是非常个人化的事情。我的意思是,你更了解自己,那么你的助手肯定和我的助手不同。所以这是相互的进步。
Yeah, I definitely agree. It's a very gradual process because typically I would recommend starting simple: just give it a task. And until today I'm still testing the limit of it. Basically I'm trying everything I touch my keyboard and mouse and try to dedicate it to it and see how it works. So you almost have to overstretch it to see where the limit is. And also I would describe this as building a relationship between you and the bot, because it's very intentional design that all the capabilities are shared—the tools, the connectors, the skills—but the memory and preferences are local to each individual bot. So it's kind of like abstracting those reusable components into the role to help people manage and then form this relationship. It takes a bit of time, but try it: start with a simple task, then start a recurring task, and then set up routines, let it automate it, do it, and maybe it burns too much token and tone down a bit, and then get connected to more tools, skills, customize it, so it's like gradually it gets more mature and it knows you. So it's even a very personal thing. I mean, you know yourself better, and then your assistant definitely would be different from my assistant. So it's this mutual progress.
这就像入职新员工,对吧?随着时间的推移,你可以更信任他们。
It's like onboarding an employee, right? You can kind of trust them more over time.
是的,我觉得是这样。我觉得这是一个很好的说法。就像雇佣虚拟员工,对吧?然后你希望能够信任他们。所以,就像你在让新员工自主之前必须培训他们一样。我认为如果你设置好了,那么你实际上可以进入一个非常好的流程,你作为某种 CEO 现在可以委派任务,走开,做更高层次的战略事情。
Yeah, I feel like that. I feel like that's a very good way to put it. It's like hiring virtual employees, right? And then you want to be able to trust them. So, the same way like you would have to train a new employee before you let them be autonomous. I think if you set that up then you can actually get into a really nice flow where you as the sort of CEO can now delegate and step away and do like higher level strategic things.
现在我们现在都可以成为产品的 CEO 了。
Now we can all be the CEO of the product now.
是的,没错。
Yeah, exactly.
酷,非常感谢两位详细介绍了你们的设置和其他一切。我猜 Lauren,人们可以在 poteto 上找到你,那是你的用户名吗?
Cool, well thanks so much both for walking through your setup and everything else. I guess Lauren people can find you on poteto is that the handle?
是的,你可以在 X 上找到我,呃 poteto,拼写是 e p o t e t o。
Yeah, you can find me on X, uh poteto with an e, p-o-t-e-t-o.
还有 Pen,你想保持匿名还是分享你的用户名?
And Pen, do you want to stay anonymous or do you want to share your handle?
哦,我也活跃在 X 上。我的用户名是我的名字 P N G Z H N G,然后加下划线。
Oh, I'm also active on X. My handle is my name P N G Z H N G and then underscore.
好的,各位。非常感谢你们的时间,Grapa 是一个很棒的产品。恭喜你们打造了它,是的,我很期待看到接下来会发生什么。
All right, guys. Well, thanks so much for your time and Grapa is an amazing product. Congrats on building it and yeah, I'm excited to see what's next.
是的,感谢邀请我们。
Yeah, thanks for having us.
是的,感谢邀请我们。真的很有趣。
Yeah, thanks for having us. It was really fun.