xAI Grok Bot Team Uses Agent Swarms to Ship Unreviewed Code

xAI Grok Bot团队利用智能体集群交付免审代码 只有当操作人员将智能体视为新员工、在将其封装为无人值守技能前手动验证执行闭环，自主AI委托才能取得成功。 Peng Zheng与Lauren Tan详细拆解了他们如何利用持久化、专业化的AI智能体集群，来自动化工作与生活中的日常工作流。Peng展示了他的幕僚长（Chief of Staff）、设计师与写手Bot如何处理各种任务，涵盖在Facebook Marketplace上挂售器材、更新Notion库存，到将单个Figma关键帧拓展为完整用户流程，以及向个人网站发布动态3D黏土微缩模型。Lauren则展示了她仿照高工艺“米其林厨房”而非软件工厂构建的分层工程架构；在这一体系中，专业Bot负责监督临时云实例，在虚拟机中运行自动化测试评估，并在人工审查前自主合并Pull Request。两人共同梳理了一套建立智能体信任的入职框架：从手动提示与纠错闭环起步，将验证有效的工作流提炼为可复用技能，最终在保留人在回路安全防护的前提下调度自主例行任务。 Grok Bot实现了职责分离的持久化、拟人化多智能体集群（例如幕僚长、工程主管和专业Bot），负责处理涵盖库存物流、行程预订和代码部署等多种工作流。 (0:00) 工程Bot的层级体系运作方式类似于“米其林厨房”而非大规模生产工厂：主管Bot负责编排架构而不直接编写代码，动态生成云智能体并在合并前自主验证PR。 (25:22) 实现真正的Bot自主性遵循渐进式信任阶梯：用户先手动迭代提示词，将成功的工作流固化为可复用技能（如`/poteto-mode`），最终将其转化为主动的、事件触发的后台例行程序。 (40:28) 该平台将全局工具、MCP连接器以及技能插件（可直接在Cursor等工具间互操作）与智能体身份解耦，使记忆与偏好严格隔离在各个具名角色之内。 (30:48)

Grok Bot creators Peng Zheng and Lauren Tan outline a shift from single-prompt chat windows to persistent, multi-agent bot architectures that execute complex engineering, design, and personal tasks. They argue that scalable automation relies on treating specialized agents as virtual employees whose autonomy expands only as users verify their reliability. Hierarchical Orchestration and Specialized Bot Swarms (25:22) Rather than relying on a single generalist model, both Zheng and Tan organize their workflows through hierarchical delegation. In Zheng's setup, a central Chief of Staff bot handles triage and routes incoming digital prompts to domain specialists like a Designer bot or a Writer bot. Tan applies a similar structure to engineering tasks, delegating instructions through a Chief of Staff to an Engineering Lead bot named Matcha, which in turn oversees several worker bots. Tan enforces a strict non-execution rule on the lead bot, restricting its role to project decomposition and supervision. Under this hierarchy, worker bots deploy ephemeral cloud agents across varying models, matching model reasoning power to task complexity. To keep architectures organized, Tan uses a meta-agent called Dr. Eggbot to generate new bots when existing personas accumulate conflicting contexts, while also auditing scheduled background routines to eliminate wasteful, high-frequency wake-ups that burn unnecessary compute. Persistent Personas and Cross-Tool Skill Interoperability (6:31) Unlike standard conversational AI interfaces that treat sessions as disposable, Grok Bot maintains persistent identities, isolated memories, and universal tool access across sessions. In multifaceted projects, Zheng brings multiple bots—such as PM, design, and engineering agents—into shared group chats where they debate tradeoffs or respond to specific `@mentions`. Even across separate threads, individual bots retain their core identity and background knowledge. Workflows also span external software environments. Zheng connects his Designer bot directly to Figma via an MCP connector using a customized skill, "Peng mode," which applies his design tokens, typography, and spacing to scale single keyframes into full screen flows. Similarly, Tan's open-source `p-stack` skill library operates interchangeably across Grok Bot and Cursor, while local daemons like Crumb manage Mac disk space and interactive test suites execute inside Linux virtual machines. Real-World Execution Across Digital and Physical Tasks (1:23) Zheng and Tan push their agents beyond text generation into multi-step tool execution. Zheng's Chief of Staff bot automates physical inventory updates in Notion after ordering 3D printing filament, manages Facebook Marketplace listings by price-matching competitors and applying weekly markdown rules, and generates 3D dioramas from phone photos to publish directly to his personal website. If a task requires secure authentication, the bot presents a browser takeover screen for manual login before resuming execution. Personal operations follow similar patterns. Tan uses a dedicated assistant bot, Suki, to automate routine Amazon reorders, send SMS pings that bypass macOS "Do Not Disturb" settings during coding sessions, and place DoorDash orders. For high-stakes tasks like booking multi-city flights across London and Amsterdam through the corporate platform Navan, Tan maintains an approval gate: the agent evaluates cabin policies, schedules, and hotel proximities, but requires human confirmation prior to purchase. Building Trust: From Manual Loops to Autonomous Routines (40:28) Tan and Zheng emphasize that agentic automation cannot leap directly to unattended autonomy. Instead, users must cultivate trust across three phases: manually prompting through an entire loop, distilling the corrected workflow into a reusable skill or slash command, and finally linking that skill to automated triggers. Tan illustrates this with expense reporting, where manual verification gradually transitions to single-shot commands and eventual automated processing of incoming receipt emails. To ensure autonomous reliability in software engineering, Tan configures agent verification pipelines where Cursor cloud agents spawn competing sub-agents across multiple models to evaluate pull requests against objective playbooks before merging. Zheng characterizes this progression as moving from manual labor (Tier 1) and prompting (Tier 2) to designing system guardrails (Tier 3), enabling humans to operate as executives delegating to self-verifying systems. Zheng and Tan frame the future of AI interfaces around personified, persistent agents rather than model harnesses, arguing that human operators must act like executives onboarding new hires. However, achieving dependable hands-off execution remains contingent on upfront human investment in skill curation, verification playbooks, and continuous routine auditing.

[0:00] Peng Zheng: I will do the first 5% and I'll create like one key frame [0:00] Lauren Tan: 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.' [14:16] Lauren Tan: the importance of actually having the agent be able to verify its own work. [17:59] Lauren Tan: I guess it's like an Asian thing, we love food [25:22] Lauren Tan: I like to call it the Michelin kitchen because to me... when you think about a Michelin-starred kitchen, right, you think about the quality, right, and the craft that goes into making those... dishes or meals. And there's still scale involved... A restaurant might have, you know, 50 seats or 100 seats, so you need to be able to produce quality at scale. [32:35] Lauren Tan: if you say poteto three times I get like a notification. [32:35] Lauren Tan: Hey, Poteto said blah blah blah... go talk to Poteto and figure it out, and come back to me with a design proposal for how we fix this issue. [35:54] Peng Zheng: my mom would never know what is a SOUL.md. [38:59] Peter Yang: I'm just going to go to bed and let this bot do everything [42:08] Peng Zheng: almost have to overstretch it to see where the limit is.

[0:00] I will have my PM design like [music] [0:02] three individual bots edited them into a [0:05] single group chat and then they'll start [0:06] to discuss and ideulate about the sync. [0:09] I [music] will do the first 5% and I'll [0:11] create like one key frame and then ask [0:13] it to scale [music] it into the whole [0:15] flow and make it like an end to end [0:17] range. [0:17] >> It lets me orchestrate like really [0:19] massive agent swarms. Sometimes I [0:21] actually don't even look at the PR until [0:23] after it's landed and then I'm like, [0:24] "Oh, okay. Yeah, that looks good." I [0:26] [music] think it ultimately comes back [0:27] to trust, right? How much trust are you [0:29] willing to give your agent or your bot [0:31] to be able [music] to do a task for you? [0:34] And I think one thing I always try to [0:36] encourage people to do is to actually [0:38] spend the time to sort of observe the [0:40] bot or agent doing the work. [0:44] All right. Hey everyone. My guests today [0:46] are Pen and Lauren who are members of [0:48] the design and technical staff on the [0:50] Grockbot team. really excited to see [0:52] their Grockbot setups for work and [0:54] personal tasks and also how to use [0:56] Grockbot to design and build the actual [0:58] product and also the most surprising [1:00] bots that they see the community making. [1:02] Welcome both. [1:03] >> Thanks for having us. [1:04] >> Thank you. [1:04] >> Very excited. [1:05] >> Yeah, very excited. I I think Grockbot [1:07] is very exciting. I think uh it's the [1:09] first persistent, you know, personal [1:10] computer product that I've seen and I [1:12] use it every day myself. But why don't [1:14] we start with this? Um maybe you guys [1:17] can show us kind of your personal [1:18] Grogbot setup and a few of the bots that [1:20] you're really proud of. Maybe let's [1:22] start with you, Pen. [1:23] >> Yeah. So um I'd like to share a couple [1:25] of bots I have. So basically I'm using [1:30] Grabbot for almost everything um in the [1:33] digital world I'm having. And at a at a [1:36] high level I categorize them by work and [1:38] life. Um and then I have like those most [1:42] frequently used bots pen here. Um, chief [1:45] of staff is my go-to bot. If I don't [1:47] know which one should take the task, [1:50] then I'll send it to chief of staff. Uh, [1:52] and one thing is you can almost ask it [1:56] to do anything in the digital world. Um, [1:58] I was asking it to buy 3D printing [2:00] filament um, couple days ago and it does [2:04] not only buy it for me. I also ask it to [2:07] add it to the inventory list. So I have [2:09] this notion database to track all my [2:12] inventory different colors of the [2:13] filament and once they once they bought [2:17] it it also like update the list for me. [2:19] >> Um I just discovered like the benefit of [2:22] trying different tasks together is [2:24] really powerful because another example [2:26] is like let's say I want to sell my DJI [2:29] mic on Facebook marketplace. Um I just [2:31] send it a photo and then tell it do a [2:34] couple things. So like sell something, [2:37] list something on Facebook. Marketplace [2:38] sounds easy, but actually there are a [2:40] lot of steps like you need to check the [2:42] latest price for the model on the [2:43] official website. I wanted to find out [2:46] what's the lowest price offered in my [2:47] area, like who I'm competing with. I [2:50] wanted to learn like how I write listing [2:52] descriptions uh in the past and also [2:55] post this new listing. And then if [2:58] nobody's interested in it, automatically [3:00] lower the price by $5 every week. So you [3:03] see like it's like a series of tasks [3:05] that I dedicate in one single prompt and [3:08] it will go execute the task and it's [3:11] very interesting that it know like when [3:14] it should write the description it [3:17] consult my writer bot so it message the [3:19] writer bot it ask like how I wrote uh [3:24] marketplace listing description in the [3:25] past and then it will able to do the [3:27] research it'll be find out like how [3:29] other people are selling it and then [3:31] eventually pull up my list. So [3:35] in the past we may do like all those [3:37] five steps manually by ourselves but now [3:40] not only can delegate individual task to [3:42] it but also like ask it to train the [3:44] tasks for you. Um my bot is replacing my [3:48] email. So like I would do things like [3:51] hey what are the action items in my [3:53] inbox today? Uh my email is still out of [3:56] manage and like for things I I [3:59] identified that can do it for me. For [4:01] example, I need to pay my friend money [4:02] for dinner. Uh I'll just ask it to do [4:04] so. And if it doesn't have access, um it [4:07] will prompt this signin page for me and [4:09] I can take over and sign in on its [4:12] computer and it will do it for me. Uh [4:14] it's replacing my calendar. So if I'm [4:17] having a coffee chat with somebody now, [4:19] like basically any edits, any event [4:21] creation, find booking like book [4:23] conference rooms, anything I just put a [4:26] screenshot and then let it do it. let it [4:28] like add the events to the calendar, [4:30] modify it, um stuff like that. Um I [4:35] noticed you have a Okay, so you have a [4:37] Okay, let's let's go up to your ping [4:38] bots. So you have a designer and a [4:41] writerbot pinned. So designer So the [4:43] writer is like your writing style and [4:45] designer is like Yeah, maybe can quickly [4:47] walk through those. [4:48] >> Yeah. So, this is a demo account that I [4:50] don't have like everything here, but [4:53] it's the exact same setup I have in my [4:55] like in my own account. Um, for like [4:58] those three are the most frequently used [5:00] like chief of staff going to manage all [5:02] other bots. And the designers like I [5:03] stayed in design tools every day and [5:05] then designer is my go-to bot. I'll have [5:08] it like brainstorm with me or like ask [5:10] it to um spin off different design [5:13] options for me. Um, and I'll just like [5:16] discuss with it and not only like work [5:18] with me on it, but also execute the [5:20] ideas. And then for the writerbots, um, [5:23] currently like I I'm not an English [5:25] native speaker. And then a lot of emails [5:28] and Slack messages, I would have like [5:30] the writerbot to polish it before I send [5:32] it. And we have the Slack and email chat [5:36] widget. So you literally see um the [5:39] draft that inline here that you can you [5:41] can review and send and decide how to [5:43] modify etc. [5:45] >> Got it. This episode is brought to you [5:47] by Granola. If you're in backtoback [5:50] meetings, you know how much work it is [5:51] to take notes live and clean them up [5:53] afterwards. That's why I love Granola, [5:56] the best AI [music] meeting notes app in [5:58] the market. Here's how I use it. Granola [6:00] automatically takes notes during a [6:01] meeting and I can add my own notes, too. [6:04] After the meeting ends, I use a Granola [6:06] recipe to extract clear takeaways and [6:08] next steps in the exact format that I [6:10] want. Then I can just share her notes [6:11] directly in Slack with my [music] [6:13] colleagues or even get Granola to share [6:15] her notes automatically. Honestly, of [6:17] all the AI apps that I use, Granola is [6:19] the one that saves me the most time. Try [6:21] it now at granola.ai/pater [6:24] and use the code Peter to sign up and [6:27] get three months free. That's [6:29] granola.ai/peter. [6:31] Now, back to our episode. You also have [6:33] a channel like me now app or something [6:35] where you have a bunch of bots talking [6:36] to each other. [6:37] >> Yeah. So [laughter] what is that? [6:41] >> Once you have bigger tasks that um you [6:45] want a different bots with different [6:47] rows have their own perspective and [6:49] contribute to it. [6:50] >> Um this is an example like if let's say [6:53] we are building an app that can [6:54] translate cat meow to English and I want [6:57] to start [6:59] review this idea from different [7:00] perspectives. So I will have my [7:05] um PM design like three individual bots [7:09] addit them into a single group chat and [7:12] I will throw this prompt and then [7:14] they'll start to discuss and ideulate [7:16] about this thing and it's interesting to [7:18] see like how they think about the idea [7:20] how they debate with each other. Um if [7:22] you want a specific bot to reply to you [7:24] just simply at mention it and only that [7:27] bot will reply to you. uh if you just [7:29] stand a message in the group chat just [7:30] like how you message your friends in [7:32] group chat they will review it and then [7:34] they will reply they will decide if they [7:36] should reply uh from their own end go [7:38] >> yeah this is really interesting I don't [7:40] think this is easy to do in some of the [7:41] other AI apps right because like keeping [7:44] the context clean and having all these [7:45] bots that have their own context chat [7:47] with each other is like pretty unique [7:48] >> yeah it's um so there's like two angles [7:52] you can think about the problem like [7:54] some people manage their tasks and jobs [7:58] by individual sessions, task and [8:00] projects. [8:01] >> Um, and we are creating bots as more [8:04] persistent named um, agents that have [8:07] their own identity, memory system, had [8:10] their own tools. U, it's really like [8:12] abstract what's reusable into the row [8:16] and then you can have them work together [8:18] by creating separate chats and group [8:21] chats for them. like even I have [8:23] multiple projects with the same designer [8:25] bot I can create another chat with it [8:28] like for a different project right and [8:31] then I can do this repeatedly for [8:33] different projects I'm managing because [8:35] one thing we noticed is a lot of time [8:38] uh in AI agent product those chat are [8:41] very ephemeral they're almost disposable [8:43] like you don't care like maybe you only [8:46] care like top three to five that you [8:48] actively working on and then for the [8:50] chats like you happened two weeks ago [8:52] like you never go back. So I think this [8:56] is a great way to like balance the two [8:58] that for whatever reusable the [9:00] capability the memory then you kind of [9:02] abstract them into different rows and [9:04] then for individual projects that has [9:08] the end has the deadline and then you [9:10] can create this dedicated sessions and [9:11] work with your bots in it [9:13] >> and but but just like a human the [9:14] designer bot like if I talk to designer [9:16] bot for mail app and bark app it knows [9:19] about both right in its context. Yes. [9:21] >> Or is it separate? [9:22] >> Yes. So when a bot is working in a group [9:24] chat, you will able to see like the work [9:27] status has changed in its own DM. So [9:29] it's the same entity. It's the same bot [9:31] working on your task. [9:33] >> Got it. [9:33] >> Yeah. And then I have other use cases I [9:35] think is pretty fun like grabbot is [9:38] replacing my podcast tool right now [9:40] because I listen to a lot of podcast but [9:42] sometimes when I read an article I want [9:44] to listen to it when I'm driving or [9:46] taking shower. So I'll just simply send [9:48] this link to my podcast bot and it will [9:52] translate it into a podcast uh that I [9:55] can listen anytime [9:58] I just post. [9:58] >> So it's very it's very um interesting [10:01] that you see it become like automate [10:03] auto ultimate translator for different [10:06] file formats. [10:07] >> Um and also I like to share this [10:10] personal project like the pipeline I set [10:13] up. Um, the story was I took a photo of [10:18] a produced shop in San Francisco, [10:21] Chinatown like two weeks ago and when I [10:24] was waiting for the bus and then I [10:26] turned it into like a 3D clay miniature [10:30] uh using AI and I thought, "Oh, this was [10:33] fun." and I put it onto my personal [10:35] website and it was very minimal website [10:38] and I put it here and then I wrote the [10:40] line that lasting at a JBC market in [10:42] Chinatown and then I thought oh this [10:45] could become a checkin app because for a [10:48] personal website like you put it up [10:50] there and people visit it once and then [10:52] nobody came back again like I don't even [10:54] visit it that often so I had this idea [10:57] of making it an internet check-in app uh [10:59] so every time I check in it can just [11:02] automatically [11:03] post the location and post the shop or [11:06] store whatever um up there. So I started [11:09] to talk to Grogbot about this idea and [11:13] eventually I was able to land it on this [11:15] pipeline that if I texted it uh a image [11:19] of the place I've been and also like who [11:21] am I with the social handle of the [11:24] person it will be able to generate um [11:27] two images in day and light mode and [11:31] also remove the background um and then [11:35] eventually post it to my website um [11:37] something like is so [11:39] >> Oh, nice. Okay. [11:41] >> It was it was uh it was very [laughter] [11:44] whimsical, fun, like personal project. [11:47] Um but I think it's a different uh [11:51] perspective to think about it is now I'm [11:53] using Grubbot as like the main interface [11:56] to to to host things instead of like [12:00] having build a portal or CMS system and [12:03] manually upload it in the in the back [12:04] end. And a lot of it is baked in [12:08] >> um [12:09] >> baked in baked in this pipeline that I [12:11] have this bot just go fetch the latest [12:13] playbook on this private repo uh and [12:16] then follow the step and then like [12:18] generate the image clean it up and then [12:20] post it. [12:21] >> Okay. So, so people can see all the [12:22] places you've been to and and the [12:25] typical designer you think about both [12:26] day and light mode. I would not think [12:27] about that. Yeah, that was good. Yeah. [12:31] And [12:31] >> yeah, so that's example that you can use [12:33] grabbot in a creative way almost think [12:36] it as like a building module in the [12:38] thing you want to build and also the [12:40] best part is it can guide you step by [12:42] step towards that goal. [12:43] >> And before we move on to Lauren um your [12:45] designer bot so are are you still like [12:47] drawing boxes in Figma or you tell your [12:49] bot to do everything and you just re [12:50] review? [12:51] >> We so as xaii we do we we design in [12:55] whatever tools we need. Uh we use Figma, [12:58] we use canvas based tools, we have our [13:00] own prototype environments and sometimes [13:02] we also submit our own PR. [13:04] >> Um so the designer bot is really helpful [13:08] to bounce in ideas with me. I also make [13:10] quick changes um because it connect the [13:13] MCP Figma MCP through the connector. So [13:16] it it will be able to like directly [13:18] >> change frames and update the design and [13:21] use design token. Um I have this skill [13:24] that the designer bot is able to use [13:26] that so it know like how my Figma uh [13:30] file setup is what's my design system [13:33] and when to use the right like colors [13:35] typography spacing um when it created [13:38] design for me. [13:39] >> Okay. So maybe it can do like the first [13:41] 80% you can manually tweak stuff or [13:42] something. [13:43] >> Yeah. [13:44] >> Yes. Yeah. Or sometimes like my workflow [13:47] has become I will do the first 5%. [13:49] >> Uh I would create like I would build the [13:51] system and I'll create like one key [13:53] frame and then ask it to scale it into [13:56] the whole flow and make it like an end [13:58] to end uh screens. [14:00] >> Oh, awesome. You should like open source [14:02] your design skills like Lauren did with [14:03] her PP stack. [14:05] >> Yeah. Yeah. Interesting. Like we have [14:07] this [14:07] >> this skills follow our names like Lauren [14:10] [laughter] mode and I have a pen mode. [14:14] >> Pen mode. Yeah. Think exactly. [14:16] >> Awesome. Pen, thanks so much for sharing [14:17] all this. This is super cool. Lauren, [14:19] thanks for waiting. I would love to see [14:20] how your setup is, too. It's probably [14:21] pretty different from [14:23] >> very different. Yeah. Uh, I wanted to [14:25] share I guess more of the engineering [14:27] use cases. So, one thing I've been [14:28] recently working on is adding better [14:30] support for Omachi, uh, which is a Linux [14:34] distribution that DHH and a couple of [14:37] other really awesome open- source folks [14:40] have been working on. Um, and I've been [14:43] using Grockbot to essentially kind of [14:45] control uh this is like a this is a [14:49] special build of Omachi which runs in a [14:52] virtual machine on your Mac. [14:55] >> Uh, which is still like very alpha but [14:57] it works pretty well. [14:58] >> Uh, but it's really powerful because I [15:00] can actually get my agent to literally [15:03] like script and control the VM. So uh [15:07] what better way to build support for you [15:10] know this particular flavor of Linux [15:12] than to have graphbot actually be able [15:14] to open it on my computer uh and you [15:17] know like interactively install it like [15:20] it installed the graphbot for me. You [15:22] can see like it's actually doing that [15:23] right now. [15:23] >> Okay. So graphbot is on its own computer [15:25] but it's controlling your computer [15:27] through omachi. So yeah, so Grapa has [15:30] its own computer. Uh but it also has a [15:32] local execution demon I guess that runs [15:36] on your computer. So if you if you tell [15:38] it to like another [clears throat] thing [15:39] I I often actually do is this I have [15:41] this agent called crumb and I often just [15:44] like run out of hard disk space. So I'll [15:47] just tell it like hey go through my hard [15:48] disk and like you know find where is all [15:50] the space going and it just goes off and [15:52] like find stuff and to delete it. Uh but [15:55] yeah, it has like full access to your [15:57] computer and uh I mean depending on the [15:59] permissions you give it, but it can uh [16:02] you know then in this case for me [16:04] because I wanted to try out Omashi and [16:06] you know fix some bugs or uh work on the [16:10] experience a little bit for Linux, it [16:12] was very powerful to be able to actually [16:13] have the bot control the virtual machine [16:18] itself, the Linux virtual machine on my [16:20] computer. [16:21] uh cuz yeah that's like one thing I [16:23] always talk about is like you know the [16:24] importance of actually uh having the [16:27] agent be able to verify its own work. So [16:30] like this is one example where the agent [16:32] can verify it work on your computer in [16:35] addition to of course like you know it [16:36] has its own computer uh that you can you [16:39] know take a look and and see what it's [16:41] doing. Uh but uh yeah that's like one of [16:45] the the main use cases I have uh is is [16:48] doing like a lot of engineering work uh [16:50] on on graph. [16:53] >> Um another thing I manage is like my own [16:56] personal stack of plugins Pstack. Um so [17:00] all of the work I do on this plug-in [17:02] basically happens in here. Um so uh you [17:07] know I'll kick off cloud agents cursor [17:09] cloud agents uh which um will um [17:14] essentially do eval for me. So in order [17:17] to like make a change to my set of [17:19] skills that I can trust that the skill [17:22] the improvements are actually going to [17:23] you know result in better performance uh [17:26] I have uh like an eval playbook in PSA [17:29] that teaches the agent how to actually [17:32] like it'll spawn a couple of sub aents [17:34] of different models. It'll try out the [17:37] prompt. Uh it will, you know, make sure [17:39] that [17:41] uh the you know the intent the goal that [17:44] we actually set out to achieve was [17:46] achieved and then only if it does [17:47] achieve it, it will actually uh you know [17:50] land that that pull request. [17:53] >> Um yeah, so that's uh quite a lot of the [17:56] engineering work that I do. [17:59] >> Okay. And and I noticed your your your [18:01] Grock bots have very different names [18:02] from Pence Grock. They're more [18:04] [laughter] [18:05] they have more they have more [18:06] personality than just, you know. [18:09] >> Yeah. I like I I I asked my So I have [18:12] this bot called Dr. Eggbot, which I've [18:14] also kind of shared uh on my socials. Uh [18:18] but Dr. Eggbot is essentially my [18:20] designer bot. Uh I mean, not a designer [18:23] bot, but like a bot designer bot who [18:26] designs bots for me. uh with all of the [18:29] rigor, engineering rigor that I expect [18:31] from the bots that I use. Um and uh I [18:35] also get it to come up with food theme [18:37] names. I don't know. I guess it's like [18:39] an Asian thing. We we we love we love [18:41] food so we got to call it, you know, [18:43] call my botsman cute like Dau Fuku Gza. [18:47] [laughter] [18:48] So that's pretty fun. [18:50] >> Um and then similar to Pang, I also have [18:53] a chief of staff. I have a personal [18:54] assistant. [18:56] um personal assistant helps me with [18:58] things like lunch, my flights, uh uh [19:02] booking hotels. Uh sometimes I I ignore [19:05] her and then she gets a little bit [19:06] sassy. [laughter] [19:08] >> Uh then my chief of staff uh yeah like [19:12] helps me with just all sorts of random [19:14] things. But uh if you you see in my [19:17] sidebar, I have a bunch of like engineer [19:19] bots here. I have some social bots as [19:22] well that manage my you know my social [19:24] media. [19:26] >> Um so a lot of the work that I do goes [19:28] through these two or these three agents [19:31] I guess like either I'm you know talking [19:34] to the chief of staff and my chief of [19:36] staff isn't working with like my in lead [19:39] uh matcha here uh or you know if it's [19:42] Suki my personal assistant like she [19:44] might you know um book flights for me. I [19:48] haven't set up like a personal assistant [19:49] team yet because I haven't I don't have [19:51] that many I'm not that busy I guess that [19:55] I need like you know a team of bots [19:57] helping me there. So so far she's been [19:59] she's been doing a really great job. [20:01] >> Uh and then Dr. just comes up with bots [20:03] for me whenever I'm like I'm starting to [20:06] realize like oh you know I'm starting to [20:08] use like one agent for a lot of things [20:10] and that's kind of um [20:14] uh I'll say like this means I'm adding [20:17] like a lot of context to that agent [20:20] that's [20:20] >> you know kind of very diverse and very [20:23] different [20:24] >> which can be like maybe confusing. So uh [20:27] one of the things that Dr. Dr. Eggbot [20:29] does as well. It's like has a routine uh [20:32] where [20:33] uh yeah like here where it will so this [20:36] is like a routine that Dr. Eggbot will [20:38] set up for you where it'll look through [20:40] all of the bots that you have um and the [20:43] transcripts that you've uh you know [20:45] you've been chatting with them about uh [20:48] and it'll propose things like [20:50] >> uh you know this bot keeps [20:53] >> keeps doing this thing that's not so [20:54] good and then it'll suggest like some [20:56] improvements whether it's skills or a [20:59] new bot or uh adjustments to routines. [21:03] Uh it also goes through your existing [21:05] routines as well and make sure that um [21:07] they're efficient because routines are [21:09] actually quite uh if you have them run [21:12] too often it can be quite expensive on [21:15] your usage because it just keeps you [21:16] know waking up the agent like if you do [21:18] like I don't know every 10 minutes [21:20] that's like a lot of wakeups. Uh [21:23] >> so um Dr. Abot does help a lot with [21:25] that. [21:27] >> Okay. Well okay I'm going to give you a [21:28] couple questions. So [21:30] >> sure [21:30] >> for for Sukiebot. So you actually rely [21:33] on to book flights and stuff like do you [21:34] actually rely on to make like big [21:35] purchases or what like what what's the [21:38] most autonomy you've given it to? [21:41] So uh yeah, I've actually booked my so [21:44] I'm going to give a talk soon uh in [21:47] London and Amsterdam as part of the [21:49] compile on the road conference and so uh [21:53] hopefully I don't leak my flight details [21:55] but uh yeah like basically I booked the [21:58] entire trip through Tuki um and this was [22:01] actually quite I was actually quite [22:03] impressed by how she did it because it [22:05] was a very it was kind of a complex like [22:08] multi-step flight. I don't know what you [22:10] call it. Like not it's not a simple just [22:11] like return flight from Orange County to [22:14] London. It was like Orange County and [22:17] then to like Denver and then to London [22:20] and then London to Amsterdam and then [22:22] Amsterdam back to back to my home. [22:24] >> Yeah. [22:24] >> Uh and I booked it entirely through uh [22:27] like we we use this uh what's it what's [22:31] called Navan, I guess. [22:33] >> Okay. Corporate thing. Yeah. [22:34] >> Yeah. It's like a corporate thing to [22:35] like book flights for you. Uh but it did [22:37] all of the work in in finding the best [22:39] you know legs and the best uh the best [22:42] class of flight that I could get away [22:44] with uh the best timing you know and it [22:47] also figured out all the hotels for me [22:49] uh which were in close proximity to the [22:51] event. [22:52] >> So that was like a really it did it all [22:55] entirely autonomously. Uh I just I I [22:57] just gave it a link to a Slack message [22:59] and with some of the details about the [23:02] conference and said book some hotels and [23:04] flights for me and you know check in [23:06] with me before you hit approve uh sorry [23:08] if you hit book and then it came up with [23:11] some proposals. [23:12] >> Uh and I just yeah I just hit yes and it [23:15] just booked everything for me. So that [23:17] was very very nice. It's probably the [23:19] biggest yeah biggest purchase I've done [23:22] through um through GOP [23:25] >> I guess. Uh, okay. But but you still [23:27] kind of got to like outpoint chat what [23:29] he plans to book and stuff, right? [23:30] Before you [23:31] >> Yeah, it wasn't uh Yeah, it wasn't to [23:33] the point yet where, you know, I I guess [23:35] I didn't want to have like some crazy [23:38] redeye flight uh or or, you know, like [23:40] be put in like the back of the plane. Uh [23:44] so I Yeah, I I I wanted to approve it [23:46] first. Yeah, I have a very similar bot [23:48] uh that mons I'm I'm trying to book a [23:50] vacation to Japan again and like you [23:52] know you can set up like price alerts on [23:53] Google flights but like the bot's more [23:55] smart because it can find like other [23:56] routes and stuff that I haven't set up. [23:58] So it actually saved me like a bunch of [24:00] money recently. [24:01] >> Yeah. [24:01] >> Yeah, it definitely can. Yeah, [snorts] [24:04] I also use it to book uh buy my uh [24:07] personal things like I we're running out [24:10] toilet paper for example. I'll just [24:11] tell, hey, [24:12] >> get us reorder our Amazon order for for [24:16] toilet paper and it just goes off and [24:18] does that. There was a point actually [24:20] like very early in the grabbot [24:22] development where we were still kind of [24:25] uh iterating on it that uh one of the [24:27] topics that came up was your bot being [24:30] able to text you. So, one of the things [24:32] I actually added to Tuki a while back [24:34] was this uh I basically gave her her own [24:36] phone number through uh a third party [24:39] service. Um, and then she would send so [24:42] then she could actually text me because [24:44] one thing I'm really bad about is [24:45] remembering to eat lunch. [24:46] >> So I need the reminder uh and and then I [24:49] often will sometimes just get I'll just [24:51] sometimes have I don't know like do not [24:52] disturb mode on on my on my Mac. Uh and [24:55] I don't see the you know the the ping uh [24:58] but then I get the text message and it [25:00] makes a sound and like oh yeah I have to [25:02] order lunch. Uh, and so I also taught [25:05] uh, Tuki how to use Door Dash, of [25:07] course, and you know, get me get me [25:09] lunch. Uh, so I don't have to to worry [25:12] about it so much. [25:13] >> That's funny. Lunch is the highlight of [25:14] my day, so I always look forward to [25:16] [laughter] never forget my lunch. [25:17] >> Yeah. Yeah. Some I don't know. You just [25:19] get too in the zone, you know, and then [25:20] you just forget to eat. [25:22] >> Okay. So my second question is I noticed [25:23] you have a lead and two engineer bots, [25:26] right? So basically you [25:28] >> I have three. each each one works on a [25:30] ticket or something or how did that [25:31] dedicate responsibility to [25:33] >> um yeah so I actually mostly talk to my [25:35] chief of staff and then my chief of [25:36] staff talks to the engine lead uh but I [25:39] set these bots these engineer bots up [25:41] with Dr. Eggbot of course uh but one of [25:44] the things I kind of drove uh I I told [25:47] Dr. for eggat to do is to make sure that [25:50] the end lead's job is really to you know [25:53] break down tasks into smaller pieces and [25:56] delegate and uh supervise other bots [25:59] rather than do work on its own. [26:01] >> So you can see in the the description [26:03] that basically it has like even the the [26:07] bot ID is in there uh but it tells it to [26:10] you know never do work on your own and [26:12] always uh delegate to those four bots. [26:16] Uh and then it will it basically does [26:18] that. Um and uh a lot of the work as [26:21] well is done through cloud agents. Um so [26:26] uh for example [26:29] uh one of these bots will uh in order to [26:31] do the work they will like create a [26:33] cloud agent uh and then you can you know [26:36] open that in cursor or on the web. Um, [26:40] >> okay. [26:41] >> But yeah, this allows me to essentially [26:42] make use of different models depending [26:45] on the type of task. Uh, cuz you know, [26:47] some tasks are are more straightforward [26:49] and I just want like really fast uh [26:52] speed. Uh, other tasks, you know, maybe [26:54] require a bit more planning and a bit [26:56] more reasoning and then I'll might [26:58] adjust the reasoning level. So, cloud [27:01] agents in graphbot give me a lot of like [27:03] very granular control I guess. Um and [27:06] also ensures that you know like my bot [27:08] kind of stays I I kind of treat each of [27:10] these engineers as more like like their [27:12] like their supervisors in a way. [27:15] >> Uh we have like so we have like the lead [27:17] that manages them but then they [27:19] themselves try not to do work on their [27:21] own. Uh so they always spawn um cloud [27:24] agents for me. So this kind orchestrate [27:27] yeah they they do all the message [27:29] passing but lets me orchestrate like [27:30] really massive agent swarms [27:33] >> uh to do like you know like if I need to [27:35] tackle a very big project I can just [27:37] talk to my chief of staff and say you [27:39] know break this down you know I have a [27:41] giant project uh maybe let's do some [27:43] planning first in a notion doc and then [27:45] it will start thinking about the phases [27:47] of the of the work that it needs to do [27:50] and then it'll delegate it all to the [27:52] engine and who will then break it down [27:54] even further and have the the bots kind [27:57] of work on that. [27:58] >> Okay. Then each one will come back with [27:59] like a PR and stuff and the bots will re [28:00] review it first. Is that kind of how it [28:02] works? [28:02] >> Uh yeah, essentially. Uh although for [28:05] me, yeah. Uh one of the things that I've [28:07] done a lot uh that I talk about a lot as [28:09] well on X is the idea that like for [28:13] example the Grothbot codebase, we've [28:15] tried to make it very how do I say uh [28:19] make it very agent friendly. The [28:21] architecture and the codebase is very [28:23] agent friendly. So, a lot of times it [28:25] actually just lets me automerge my PRs. [28:28] So, uh sounds kind of scary to say, but [28:31] some sometimes I actually don't even [28:33] look at the PR until after it's landed [28:35] and then I'm like, "Oh, okay. Yeah, that [28:37] looks good." Uh, [laughter] [28:39] >> got it. [28:40] >> Yeah. [28:41] >> Wow. Okay. So, so like you have like a [28:42] software factory. Is that is that the [28:44] buzz buzz word? [28:46] >> Uh, I like to call it the Michelin [28:48] Kitchen because I think software [28:51] factory. Yeah. Uh, yeah. I like to call [28:53] it the Michelin kitchen because to me [28:54] Michelin kitchen [28:56] >> you see like there's this uh when you [28:58] when you think about a Michelin starred [29:00] kitchen right you think about the [29:02] quality right and the craft that goes [29:04] into making those uh you know the the [29:08] product right which is a a dish or a [29:10] meal uh and there's still scale involved [29:13] right like you know a kitchen a [29:16] restaurant might have you know 50 seats [29:18] or or or 100 seats uh so you need to be [29:21] able to like produce quality at scale. [29:24] So that's the aspirational vision I have [29:26] is like I want to set a Michelin kitchen [29:28] not just like you know a factory which [29:32] uh interestingly I think a lot of [29:33] people's like when you say factory they [29:35] have this connotation like oh you know [29:36] it's just like it's massured it's like [29:39] yeah slob is like low quality [29:42] >> but I I always try to aspire to do both [29:44] right like how do we get quality at [29:46] scale so yeah this is uh a lot of the [29:49] work I've done on pack and like using [29:51] rockbot together with it is in that line [29:53] of you know how do we create that [29:55] Michelin kitchen? Yeah, you should get [29:57] that you should make like a concurge bar [29:58] or something that talks like you know, [30:00] hey, we curated this PR from the local [30:02] farm and uh [laughter] [30:03] talk like a talk like a vintage [30:06] >> Yeah. Yeah. Yeah. We need like artisal. [30:07] Here's an artisal PR for you to to [30:09] >> review. So, so basically like Pun's job [30:12] is just to like take your open source [30:14] skills and somehow build into the [30:15] product. Is that kind of what PP is [30:16] trying to do or [laughter] No, I'm just [30:18] kidding. It's like Yeah. Yeah. No, it's [30:21] great. You have you definitely have a to [30:22] system set up for [30:24] >> Yeah. I think the interesting thing for [30:26] me as well is like I found that with PAC [30:28] and with like some of the so internally [30:32] we uh we have like everyone has like not [30:34] everyone but a lot of people have their [30:36] own mode skills. So like potato mode uh [30:39] is is is what is called uh in open [30:41] source but then internally it's called [30:43] lauren mode and you know like ping was [30:45] sharing there's ping mode and like a lot [30:48] of different you know engineers and you [30:49] know designers PMs have their own mode [30:51] skills um and I think this allowed us to [30:54] really compound on some of the skills [30:58] that we've already set up in the [30:59] codebase. So yeah big a big motivation [31:02] for uh uh of of the you know redesigning [31:07] and architecting the graphbot codebase [31:10] was so that everyone right could [31:12] contribute at a high level not just [31:15] engineers. So a lot of care has been [31:17] taken into doing that. Um and of course [31:21] yeah like people are kicking kicking [31:23] these agents off in graphbot um as well [31:26] as sometimes in cursor. Uh and the nice [31:29] thing about cursor and graphbot plugins [31:31] is that they interoperate. So you can [31:34] install PAC for example on both uh [31:36] grabbot as well as cursor and kind of [31:39] reuse the the same skills. [31:41] >> Interesting. Yeah, I didn't think about [31:42] this. So So your bots can also call [31:44] different skills, right? Like let's say [31:45] I have like a right like me scale or [31:46] some sort of scale. You just trigger it [31:48] the slash. [31:49] >> You can Yeah, if you do like slash slash [31:51] potato mode, you'll see that um it's uh [31:55] you can use that skill there. Um, and [31:57] you can always look at the marketplace [31:58] as well. Um, I actually don't know if [32:01] there are other plugins that are just [32:03] skills, but PAC is one of them. Uh, but [32:06] a lot of the plugins we have also come [32:08] with skills. So, uh, in addition to [32:11] like, you know, an MCP server. So, for [32:13] example, like if you I don't know if you [32:14] install the the X uh the X plugin for [32:17] example, it includes I think one skill [32:19] in addition to the tool. Um, and yeah, [32:23] like so [32:24] >> plugins can contain like lots of [32:26] different pieces of functionality that [32:29] extend your grabbot instance um to do [32:33] like lots of really interesting things. [32:35] >> Awesome. Uh, before we wrap up your [32:37] share, like um I'm I'm curious about [32:39] your social bots like um [32:41] >> yeah, [32:41] >> how how do you separate the signal from [32:43] the slop [32:44] >> on X? [32:46] >> Well, yeah. So, a lot of my social bots [32:48] are connected to uh X as well as like [32:51] LinkedIn uh which I guess I I tend to [32:54] use X more than LinkedIn. Um but I have [32:57] u uh I have like one bot called Potato [33:01] of course who goes through all of the [33:03] all of my mentions. So like uh I have [33:05] this like meme on X where if you say [33:08] potato three times I get like a [33:10] notification. [laughter] Uh and that [33:12] that that's this bot basically. this bot [33:15] is like has a routine where it's [33:17] checking x uh I think I set it to every [33:19] 30 minutes or so uh to check for all the [33:22] mentions uh and then I just jump in [33:25] there and I'll often respond uh or it'll [33:28] just kind of like aggregate like you [33:29] know like you know five users have [33:30] reported the same bug right and maybe [33:32] you should take a look at that and so I [33:35] I kind of think of the socials as not [33:37] just [33:38] not just managing my I guess my social [33:42] uh networks but also as a form of like [33:44] my outer loop where since I'm using the [33:47] socials a bit a bit for work as well. Uh [33:49] it means that you know I can stay on top [33:51] of like the pulse of you know people are [33:54] reporting bugs or they're running into [33:56] issues or you know maybe uh a new [33:59] feature that we launched is confusing [34:01] you know like I want to be able to [34:03] >> uh like stay connected to that and have [34:06] that feed my inner loop of you know my [34:08] engineers cuz then because these are all [34:11] just bots right I can I can always say [34:12] like I can go to like matcha and say [34:14] like like hey uh you know uh potato said [34:18] blah blah blah [34:19] Uh, and then the context can just be [34:21] grabbed there, right? I don't have to [34:23] repeat all of the X context that I got. [34:25] I don't have to copy paste anything. I [34:27] just say go talk to potato and figure it [34:30] out, right? And come back to me with a [34:32] design proposal for how we fix this [34:34] issue. Mh. [34:36] >> So I think like yeah when you for me the [34:38] the the the power of grapa comes when [34:41] you connect all of these different [34:43] things together and have them kind of uh [34:46] I don't know if there's a better word [34:47] for like but synergize you know like the [34:50] the context can can be grabbed from lots [34:52] of different places um and kind of be [34:56] grouped together in a way that is much [34:59] richer than if I were to type it all out [35:00] by myself. [35:01] >> Yeah. Yeah. I mean like context is [35:03] everything with this AI stuff and um [35:04] Yeah. [35:05] >> Yeah. Basically all all these bots live [35:06] in the same house and they can talk to [35:07] each other and like you know pass [35:09] conscious to each other, right? [35:10] >> Yeah. [35:11] >> Cool. Uh and uh where where does potato [35:15] come from? Do you just really like [35:16] potatoes? [laughter] [35:18] >> Uh yeah, it was just like a I don't even [35:20] know. I think it was like when uh this [35:22] was like many years ago and I was just [35:24] playing like an M MMO and I was like I [35:26] couldn't come up with a name and I just [35:28] I I have a thing for food as you can [35:30] tell from all my bot names. [35:32] >> So I was like oh potato why not and I [35:35] because I couldn't be I wasn't really [35:37] thinking that hard and then potato with [35:38] the normal spelling was taken. Mhm. [35:41] >> So I was like, "Oh, okay." I was I was [35:42] actually learning Japanese at the time. [35:43] So I was like, "Oh, okay. Maybe you [35:44] spell it like potato, like in like the [35:47] Japanese katakana." And so and that and [35:50] that kind of stuck ever since. [35:52] >> Got it. Got it. Okay. Cool. All right. [35:54] Well, this is super helpful. Yeah. And I [35:55] I just want to ask a few wrap-up [35:57] questions. So I I I feel like when I [35:59] look at the Grockbot interface, it's [36:01] almost like a version It's kind of like [36:03] a version of Slack where you have your [36:04] DMs, your teammates, right? And then you [36:06] can add them to your group chat and [36:07] stuff. And and you don't have to share [36:09] any secrets, but like there's been a lot [36:11] of stuff online about how like the stuff [36:12] is still very single player AI. It's [36:14] like one human talking to a bunch of [36:15] bots. Maybe you can talk a little bit [36:16] about like how you can make the stuff [36:18] like so like you and so Lauren and Pen [36:20] can also talk to a bunch of bots [36:21] together or or something like is there [36:24] is there any interest in doing that or [36:27] ping you want to talk about it? Yeah, I [36:28] think uh when we started designing [36:30] Grabbot, we started from a mindset of um [36:33] how can we enable people to graduate [36:36] from manually manage the task themselves [36:38] because personally I see it like three [36:40] different tiers of working like in the [36:42] past we work in a context ourselves. We [36:45] draw rectangles in Figma, we typing text [36:47] in ocean, we sending messages in Slack [36:49] and now people already got into a stage [36:51] that we are like prompting AI to do that [36:55] for us. Um but in a lot of setup um that [37:00] you've seen people are doing right now [37:01] like I believe like we are moving [37:03] towards in the future that um instead of [37:05] like manually managing all those bots uh [37:08] people are started designing the system [37:10] design the guardrail um how do you make [37:12] the bots work together um with other [37:15] human um with you um autonomously that [37:18] was that agency and intelligence [37:21] ingested and a lot of AI agent products [37:24] was very model ccentric uh because you [37:27] can see from the UI it's very model [37:28] oriented uh we built harness around the [37:31] model and then we build product around [37:33] those harness so the interface was like [37:35] tailored for like along the way we [37:37] created a lot of technical jargon and [37:39] concepts that like my mom would never [37:42] know what is a soldout MD so um it's an [37:45] experiment to like see how can we make [37:48] the concept the system model match this [37:50] user's mental model and then we found [37:52] this like personified named agents [37:55] worked pretty well and then it resonate [37:57] with people because that's how you [37:58] operate with the physical world today. [38:00] Like you don't need to know how the car [38:03] insurance company work but you know who [38:05] to call like you you have an agent for [38:06] that. If you have a car accident you [38:08] don't need to know how the medical [38:09] system is working but as soon as you [38:12] have uh health insurance you have this [38:14] doctor's office all those infrastructure [38:16] in the real world. So have that point of [38:19] contact become your like main input [38:23] output interface. um is a paradigm shift [38:27] for me because we see a like internal [38:29] like workflow become I catch a b bunch [38:34] of bugs then I just take screenshots and [38:36] annotate them and then dump into a bot [38:38] and then it will create and post it into [38:40] this notion page that my engineers bots [38:44] would pick it up from there. So it's [38:46] kind of like we started to use the app [38:47] in a very indirect way or agentic way. [38:51] Uh but somehow it was able to like [38:53] connect all the dots together and then [38:55] organically grows into like new [38:57] workflows in a very autonomous and long [38:59] running way. Maybe that's like a general [39:01] best practice overall using this stuff [39:03] like it's just like because in some ways [39:05] like the humans in a loop because we're [39:06] also passing context from one to another [39:08] and like kind of try to provide [39:10] feedback. And I I don't know. I'm always [39:11] a little bit skeptical when people say, [39:13] you know, I'm just going to go to bed [39:14] and let this bot do everything [39:16] [laughter] [39:16] and like, you know, just make a data [39:18] design or build me app and like I'll [39:19] come back and it's perfect. But I guess [39:20] maybe you guys have more experience [39:21] doing this. Like do you guys have any [39:23] tips for maybe letting these bots run [39:26] longer without going off the rails or [39:28] like kind of actually producing stuff [39:29] that you want other than using Lauren [39:31] scales? Like is is there other tips to [39:33] to do this? Yeah. [39:34] >> Uh I think the general thing for me is [39:37] Yeah. I mean you don't have to use like [39:38] PAC or anything. But I think there is [39:41] something to do with with skills. Um, [39:44] and I I talk about this a bunch, but I [39:46] think it ultimately comes back to trust, [39:48] right? Like how much trust are you [39:50] willing to give your agent or your bot [39:53] to be able to do a task for you? [39:55] >> Yeah. And I think one thing I always try [39:58] to encourage people to do, you know, [39:59] whether you are a PA stack user or not [40:02] is to actually spend the time to sort of [40:04] observe the bot or agent doing the work [40:07] and then you course correct it. But then [40:10] at the end of the the conversation, you [40:12] turn that into a skill, right? You you [40:14] you turn that into a reusable skill. So, [40:16] for example, I don't know, maybe you [40:18] have um um a skill for or you you ask [40:22] your bot to, you know, fill up an [40:24] expense report form, right? Based on go [40:26] through all of my, you know, expenses [40:28] this month and file the, you know, put [40:30] it put it all into the expense tool, [40:32] find the receipts for my email, blah [40:34] blah blah. Um, and once you've kind of [40:37] iterated a lot on that first loop, [40:39] right, if you, you know, you want to be [40:40] able to get your skill good enough to [40:42] the point where you can do like slash, [40:44] you know, expense report, right? and it [40:46] just does it well, right? Yeah. [40:48] >> And when you've got gotten to a point [40:50] where you can just do like a one shot, [40:52] right? You just say slashexpense report [40:55] and it just does the the thing for you, [40:57] >> then that actually gives you much more [40:58] confidence to walk away and say, "Okay, [41:00] now you can just go off and [41:01] automatically, you know, uh set up a [41:04] routine, right? Like every time I have a [41:05] new email come in that has an receipt, [41:08] you know, automatically file this report [41:10] for me." Um, and you don't have to [41:13] babysit it as much. But I think it yeah [41:16] there's no secret to that in my opinion [41:18] is like whether it's like doing design [41:20] work or engineering work or life admin [41:24] work it all comes back to how much trust [41:27] have you get you do you have in your bot [41:29] and if you don't trust it yet then you [41:31] know there's things like skills uh you [41:33] know prompting that kind of really help [41:35] you build that trust uh but yeah I think [41:39] you need to it has to be kind of gradual [41:41] it's very hard to get someone to you [41:42] know who's maybe never used a bot [41:45] before or they don't have a lot of [41:46] skills set up and then say, "Okay, now [41:48] just, you know, you can walk away." Of [41:50] course, they're not going to be able to [41:51] trust it, right? Because they haven't [41:53] seen it do, you know, what it can do. [41:55] But once you've done it once or twice [41:57] with the agent and you know that they [41:59] can repeat that, [42:01] >> then you can actually free yourself up [42:03] and [42:04] >> free, you know, go do other things, [42:05] right? Because now you can trust your [42:06] bot will do the right thing. [42:08] >> So, okay. So, okay. Okay, so I guess [42:09] last question is if I just download [42:11] Grockbot, then it sounds like the steps [42:13] are maybe set up a bunch of bots for [42:15] your manual workflows, go through them [42:16] manually, and then if you can do a [42:19] comfortable state, tell them to build a [42:20] scale and then eventually set up some [42:22] routines to like do it proactively. Is [42:24] that kind of what you guys recommend? [42:26] >> Uh that that's my flow. I'm kind of [42:28] curious paying about your your own [42:30] >> Yeah, I definitely agree. It's very [42:31] gradual process cuz um typically I would [42:34] recommend like start simple just give it [42:36] a task and like cuz until today I'm [42:39] still testing the limit of it. Basically [42:41] I'm trying everything I touch my [42:43] keyboard and mouse and try to dedicate [42:45] it to it and see how it works. [42:47] >> So you almost have to overstretch it to [42:50] see where the limit is. Um and also I [42:53] would describe this as like a building a [42:55] relationship between you and the bot. um [42:58] because it's very intentional design [42:59] that all the capabilities are shared, [43:01] the tools, the connectors um the skills [43:05] um but the memory and preferences are [43:08] local to each individual bot. So it's [43:11] kind of like abstract those um reusable [43:14] components into the role um to help [43:17] people manage and then like to form this [43:20] relationship. It takes a bit time but [43:22] like try it with start a simple task [43:25] then start a recurring task and then do [43:27] this set up routines let it automate it [43:30] do it and maybe it like burns too much [43:33] token and tone down a bit and then get [43:35] connected to more tools skills um [43:38] customize it so it's like gradually it [43:41] gets more mature and it knows you so [43:43] it's even a very personal thing I mean [43:47] like you knows yourself better and then [43:50] your assistant definitely would be [43:51] different from my assistant. So, it's [43:54] this like mutual um progress. [43:58] >> It's it's like onboarding employee, [43:59] right? You can kind of trust them more [44:01] over time. [44:02] >> Yeah, I feel like that. I feel like [44:03] that's a very good way to put it. It's [44:05] like kind of hiring employ virtual [44:07] employees, right? And then you want to [44:09] be able to trust them. So, the same way [44:11] like you would have to train a new [44:12] employee before you, you know, let them [44:15] be autonomous. [44:17] I think if you set that up then you can [44:19] actually get into a really nice flow [44:20] where you as the sort of CEO right you [44:23] can now delegate and step away and do [44:25] like higher level strategic things [44:27] [snorts] [44:28] >> now now we can all be the CEO of the [44:30] product now [44:30] >> yeah exactly [laughter] [44:32] >> cool well well thanks so much both for [44:35] walking through your setup and [44:36] everything else I guess Lauren people [44:37] can find you on potato is that is that [44:39] the handle [44:40] >> yeah uh you can find me on x uh potato [44:42] with a e p o t e t o [44:45] >> and pen do you want to stay anonymous or [44:47] do you want to share your handle? [44:48] >> Oh, I'm also active on X. Uh my handle [44:50] is my name P N G Z H N G and then [44:53] underscore. [44:54] >> All right, guys. Well, thanks so much [44:55] for your time and uh Grapa is an amazing [44:57] product. Congrats on building it and and [44:58] yeah, I'm excited to see what's next. [45:00] >> Yeah, thanks for having us. [45:01] >> Yeah, thanks for having us. It was [45:02] really fun.