Grok Bot For Product Best Practices

0:00 Very good. Let's kick things off. Hello everybody. I'm Kevin Neparco over from the SpaceX AAI product team. Very excited to be presenting Grockbot for product with my colleague. Yeah. Hey, I'm Roshan, also on the product team at SpaceX. have been spending a bunch of time using Grockbot and working on it behind the scenes. So excited to show you guys a little bit about how we use this every day. >> Awesome. And we are joined by Noah who will be moderating today's discussion.

0:26 Noah, do you want to introduce yourself? >> yeah, great to meet everybody. My name is Noah. I'm on our AI deployment team here. throughout the presentation, feel free to drop questions in the chat. I'll be answering them as well as like probably surfacing some topics up to you guys. Amazing. I wanted to start with a big change which I think we are probably all well aware of which is not only is product management the craft of product management changing but so too is the way in which everybody is building software and so I really like this DHH quote off the Lex Freriedman podcast recently the developers developer I think really capturing sort of this moment which is software is product management what should the product do it do it before, how should it do it, how should it look, and what are the priorities? Like these are all questions that now everybody is capable of wrestling with as we start to build products and software. everybody can be thinking like a product manager. And so very excited to talk about Grockbot and the ways that we've been using it on the SpaceX AI product team.

1:37 and so with that, wanted to share a little backdrop of why we built Grockbot. >> Yeah. Yeah. So, I thought it'd be helpful to just kick off with some general context and kind of why yeah, why exists and you know, we're thinking about this when when putting this together and a few themes sort of came up and a lot of this was inspired by our internal usage as well as what we heard from a bunch of customers over the history of building our products ourselves and I think a few themes emerged. one of those first ones is that it sort of increasingly felt that you wanted to work with agents and those agents weren't quite at home in a chat box. I think the chat box is actually really great. I use chat boxes all the time.

2:14 really enjoy them for all kinds of interactivity with LLMs and AI workflows. But for something about agents getting work done, it felt a little bit limiting and constraining. And so really wanting to think about what the right interface is and the right way to interact with an agent would be. and I also think if you kind of contextualize that to product management work, it this becomes even more salient, right? Like PMs are often trying to focus on outcomes. you're often working across multiple parts of the stack. So you're coordinating with various different functions at a company. You're like kind of a general manager. You're kind of thinking about design. You're thinking about analytics a lot. You're working with engineering.

2:52 And you'll see a little bit about that and how we've set it up in the demo later on. But really that PM work in particular seems to go beyond this Q&A format that is really native for the chat box. the other thing that was a really sort of emerging theme is that you know agents really need their own environments to get work done. we actually saw a lot of this with cloud agents early on where giving agents the ability to verify their work and verify their output results in dramatically better outcomes. and so the idea was well like what's the natural extension of that and I think it's giving agents their own environments to do the work in. And I think the more we started experimenting and playing with that, the more we saw that that could be a really powerful powerful thing too.

3:36 the last thing I wanted to mention here is that you know it gets really hard to babysit multiple agent threads. We had a lot of folks internally who had built their own sort of you know coordinator agent that was then calling out to other agents but that wasn't ever really done in a really first class way across products. And well, we think Rockbot does a pretty good job at this, but really this idea of, you know, as you're thinking about multitasking, you're switching between meeting one over here on this topic to meeting two over here on that topic, having some way to just keep track of all the the mess that that goes on in a normal workday. and I think we increasingly just started to think of ourselves as, well, what if we could, you know, work with our agents more like colleagues versus working with them more like software. And that was kind of the design brief for a lot of what Crockpot is. so to to carry on that theme for a little bit four things I wanted to call out here around agents as colleagues. and why this distinction is so important I think is because it it's a lot of the the inspiration behind the design of Grockbot and how it works and how we're thinking about the road map too. So you know we think that colleagues really focus more on tying tools together to produce outcomes. Obviously sort of connectors and MCPS and integrations have been a really strong and important part of the agent story across the industry. But I think what the sort of the exciting thing here is is making it so that those you can tie those tools together to get to work being done.

4:56 and so I think that's just a really important part of the story. I think longunning context and memory is really important about the here. Routines are really important here. And so you know thinking of first class ways to do those in the product really is really great. colleagues remember what you do like you work with them they remember what you do and so having that be a part of the solution is is really important. I think you know colleagues are independent and so making sure agents can be independent calling back this idea of agents having their own computers their own you know devices to do things on their own software to go work with. and then finally I just think getting the messaging to feel right is a really important part of this. You know when I DM Kevin on Slack I'm like rapid fire messaging and sometimes throwing in some memes in there and emojis and I think there's something about that experience just being really fun that makes it you know feel like you're working with a colleague. So, hopefully these themes make sense for those of you who had all the ones in the chat at the beginning from from having used Grockbot. hopefully you've seen some of this. but I guess that brings us to, well, what is Grockbot? And I want to simplify this after all of those words into two things. so if there's sort of two things that you're using when you're talking about Grockbot or when you're talking about it with your your friends and colleagues. The first is that it's an agent with a computer.

6:07 and I think we're really excited about the computer part of this. It sort of gives the agent its own machine, its own environment to operate in, just like I was talking about. and it really just feels like texting a colleague who has its own computer. and the second is that your team of bots really are AI teammates. They're AI teammates that you can give real work to. they finish jobs end to end. They come back when they're stuck and they function just like colleagues. And I've been pretty excited to see a lot of folks on X and other places talking about how Grockbot seems to be more of a doing agent. And so really excited for that to continue to be what we build.

6:44 >> Yeah. I think one of the things that we wanted to incorporate into Grockbot is thinking about working with sort of your best colleague or your best co-orker or teammate. what is that like? Well, they're generally very proactive. They have a bias to action. They're able to take real action based off of your guidance. They learn on the job. And so try to incorporate a lot of those principles directly into the product. >> Yeah. And I think there's some fun ways too like you can train them to do certain things and kind of teach as if you would teach you know someone you're working with.

7:11 >> Absolutely. Very cool. do you want to walk us through what the team looks like? >> Yeah, absolutely. So for the course of this demo, we're going to run through a few of the really exciting PMP specific use cases, but this really generalizes for any type of product builder. And so as we go through, be thinking about your own workflows, things that you could potentially automate, different bots that you could potentially build, and we can talk about those in the chat and Q&A at the end. one of the things that I found Grockbot to be very helpful with is keeping me focused. so I have this concept of the attention list, which is not necessarily what are my like strict priorities going into the week or the day, but more like where is my attention? And so I like to think about Grockbot as sort of sitting on top of my shoulder. being able to review all of the emails coming into my inbox, see what I'm responding to, see the different Slack threads that I'm actually engaging with, maybe the Slack threads that I need to know about but are missing and curate this attention list that can then be used in different ways. And we can talk about how that can be a filter to keep you focused. how that can be diffed against your priority list so that you can understand where your your priorities may diverge from where your time is actually going and there's a lot of really interesting ways in which you can start to use this attention list to really guide your work. The other is you know for the the blocking and tackling of new product development being able to very quickly research a bunch of a bunch of of different contexts as Roshan mentioned like being a product manager there can be context which lives in notion in Slack in Gmail and many other surfaces your support tools your CRM right and so being able to have an agent a colleague that you can work with who can help you collect that synthesize that and form formulate that into new product directions. very exciting and we'll go deep into that use case. And then of course we also need to be able to ship. We need to be able to deliver new products to market. and so we'll talk about some of the workflows and agents that we have built to help us accelerate shipping. So we're going to go through the team that we have built. This is a combination of things that Roshan and I and the product team over at SpaceX have found to be very useful for product management.

9:35 So, let's meet the team. first we have Kora. Kora is our chief of staff. you can think of Kora doing all of the chief of staff things. figuring out what is most important both for us and the team being able to reduce the distractions and block some of the things maybe autorespond to some of the questions that are coming in and helping keep us focused and organized both personally and as a team. The next is Emily. Emily is an engineering manager. Emily has been coached and taught not to actually write code like a good engineering manager but instead to manage a team of engineers.

10:21 and so these engineers are able to both learn from existing code and PRs that we have historically pushed across SpaceX AI and learn some of the best practices encode some of the skills from some of the best engineers across SpaceX AI to be able to work just like an engineer. We also have Ashley. Ashley's probably my go-to throughout the day, being able to ask questions of our data. So, Ashley has access to our data lake and our data warehouse and a bunch of different user insight and user research repositories. And so, you can ask very broad and general questions and Ashley is going to be able to build you that right chart or that right slide or that right insight to be able to answer your question. and as a PM throughout the day, many questions will come up and being able to have somebody that you can just DM and get an answer back really quickly has been exceedingly helpful.

11:21 Pete, PM Pete is our little product assistant and so will help with a lot of our RFC and product requirement doc writing can also run that research coordinate across these different agents and really help us build a high functioning product or we have our designer Pixel. Talk about Pixel. >> Yeah, I've been spending a lot of time working with Pixel actually. yeah, I think design is one of those things that's really sort of hard to get right with LLMs and with AI tools, but I actually created Pixel inspired by a recent newsletter post from Lenny.

11:58 and so there was a really great designer who who posted a design system there and kind of you know took that and and put some of the context into Crockbot and you can kind of have supplemented that as well with some other design best practices and I think it's been like the mo one of the more one of the more fun ways to do design and pixel has access to Figma of course so can get in get into Figma and help make designs and slides and also has access to our design system so that when we want to make something that you know matches the Crockbot brand or matches other parts of the SpaceXi brand there's a really strong starting point and then a really strong set of skills that Pixel's built over time to understand how to be creative and flexible and able to yeah be be a great AI designer.

12:38 >> Very good. And yes, as Roshan mentioned the PMS often wear many hats and so being able to have a colleague an agent that you can turn to for a lot of this work is very helpful. and also for people in product leadership, people who may be building a product organization, also very helpful to have a recruiter on the team. And so that is Rey. Rey can help us find amazing talent and help steward candidates through our product interview process.

13:14 One of the questions you may be asking yourself is why are there many agents? could this just be one agent? And I think we've sort of distilled this down into a few sort of helpful principles of why multiple agents or multiple colleagues really makes the most sense. I think one is just if you sort of map it to today's organizations and the ways in which a company is built you generally have specialists who are really good in a particular area. It's also very helpful for referenceability, right? So knowing who to turn to when you have a data question versus a design question and knowing that those are specialists in those areas. You also have scope memory, right? So each agent is learning on the job. You're coaching them. You're giving them feedback. and so they're able to learn different things. and then you can also run many of these agents in parallel and we can talk about, you know, how we DM with our agents and sort of this new workflow. It feels a lot more like you know just chatting with a colleague as you may in Slack or Teams.

14:13 >> Yeah, the referenceability one is huge for me because it feels like I I just need some way for my brain to stay on top of like what what everyone's working on. And so I know we talked a lot about that colleague analogy, but I think that one it it really helps for me to I've tried having just one agent and like my brain can't handle it. You know what I mean? So yeah. Yeah. I think that's huge. So >> cool. all right. Well, we want to get to maybe some exciting demos. and so we thought we would show you this team in action to build an example feature.

14:41 so we'll pan over to Grockbot here. A and just to orient you if you were if you were one of the folks who put a zero in the chat when we were gathering, meaning you have not used Grockbot before. well, this is Grockbot. we really hope you try it out. We might have some fun credits coming your way, I hear, towards the end of this. So stick around if you want to make it really easy to try it out. but essentially, Grockbot is meant to really just be simple. It's meant to be simple where you have a bunch of bots on the left here. You can pin those bots for the ones that you're talking with most and we'll you hopefully you're seeing some names here that we just talked through. So you see our chief of staff Kora. You see Ashley, our analyst, PM Pete and Emily. You see our team of engineers down here. You can also group and organize these bots, give them labels, customize them, their avatar. So really just a chance for you to personalize and make this yours. the other thing I wanted to call out while we're here is a lot of the power that these bots get come from the tools that they can connect to. And so we talked about particularly for product folks how those tools are just such an important part of our job, right? Like I'm living across, you know, 10 SAS apps each day.

15:43 and so I wanted to show that this marketplace is a really important part of this too. And so obviously you can have sort of plugins. We have categories where we've broken these down. Everything from agent orchestration tools to inbox tools, you know, MCPs that are in here. here. There's a bunch of research specific tools, documents if you're using KOD or One Drive, design tools, etc. And so, really one of the things that I encourage folks when they're getting set up for the first time is to just go through this and, you know, link all the tools that you're using, and kind of get them all connected. And once you do that set up once, sort of every agent benefits. And I think that's really that's really helpful. So, all right.

16:20 So, what do we want to show? I guess maybe, should we start with a building a feature and and see how that goes? Yeah, let's do it. So, let's start with just chatting with our chief of staff and get a feel for how you can chat with Kora here. So, let's just see what's in my inbox. >> Yeah, classic PM, classic product person work. Help me triage my day. And for the purposes of this live stream, this will be a demo environment here. And so we are asking the our colleagues here in Grockbot to hallucinate a little bit and create some fictitious data but these are going to be simulating our real workflows and the ways in which we >> this is your actual team that you're using every day. Right?

17:07 >> That's right. Y >> okay so we can see our inbox pulse. we have some questions coming in from Parker on voice. a few questions on the marketplace. and then let's see what have I been focused on this morning and this is where we can start to curate our attention lists here. And so this is going to go through Slack and Gmail and all of the different updates that I've made in notion this morning and figure out sort of where is my attention, what are the top things that I have been spending time on. So, looks like I've been spending time on local to cloud handoff, marketplace trust and a partner launch kit and a bunch of other things that I've already done this morning, which is incredible.

17:55 >> Productive morning, and it's what, 9, 10 a.m., if you haven't read it, Kevin has a really great post on X detailing how this attention list works, and I think it's really fantastic. It's sort of rather than a to-do list, it's to sort of have the bot shadow you. and I think you get a lot of really interesting insights from that and then can kind of steer these agents towards helping you get you unblocked on the things you're working on most. So that's the attentionless action right there.

18:16 Yeah. >> Very cool. And so we what we've seen is Kora has been reactive to the things that we have asked, but it is very easy to turn Kora proactive. And so we can say every hour triage my inbox and clear out anything I don't need to pay attention to. And so what's going to happen under the hood here is Kora is actually going to set up what we call a routine. You want to talk about routines?

18:46 >> Yeah. So routines are ways that you can automate agent tasks. I think you can sort of think of them as just scheduled tasks or repeated tasks, but they're natively agentic with with Grock. And so yeah, you can ask agents to create these. Sometimes if an agent sees you do this these tasks, do some task over and over again, it'll suggest or automatically create a routine. You can view and manage the routines from the UI and be able to click into them. And so maybe we could show that too. yeah, there you go. if you click I think it's the right arrow there. Oh no, that hides them. Okay, well got routines.

19:19 and so yeah, you can you can look look at routines. They're automated tasks. and what's great about those routines, they also maintain connectivity to all the tools that you connect. So we'll show an example. Oh, there we go. Yeah, we'll show an example of this a little bit later on. But for our analytics use cases, one of the things that I love is asking a my analytics bot to keep sort of hourly data or hourly tracking for new features that we launch. And that way you can sort of get a get a metrics dashboard when whenever you need on demand instead of having to go to a dashboard somewhere on the web.

19:48 So there you go. >> Very cool. and so while Kora sets that up, we are going to I think build a feature together live with our Grockbot colleagues. And so to kick things off, I had asked what are some of the top features that people have been asking for on Reddit and X? And so I think we have a pretty good list of research. P PM Pete went out did a bunch of research across social. also looked at our internal Slack discussions, our feedback channels. we also have a amazing user research database of all of the interviews that our user research team is out doing with users and customers.

20:37 And so there is a lot of feedback coming into our product team. >> Yeah. And synthesizing that I think is just its own job sometimes. One of the things that I found most useful about bot on that UXR phase is being able to generate insights from all the data, but also just like keep an eye and a pulse on what's new. And so normally I'd have to go in and reprompt or reprompt or reprompt when a new user research interview comes in or every morning with latest Slack updates. But with those routines we just showed or with some quick prompt like this, you can kind of get it all at once. So, I think it's awesome. Any cool ideas in here? You see?

21:10 >> I'm loving two-way voice. But, while we have you all here, maybe you can throw in your top Grockbot feature request for anybody, who has been using the product. Drop them in the chat. We will add them to our user research database. >> Connect Rockbot to Tesla. Account switcher. Ability to switch models. Sync of draft messages. Okay. Okay. Voice pin to top and note thread. Okay.

21:48 Two-way voice. I'm seeing a lot of votes for for two-way voice. >> Let's try it. >> Okay. So, why don't we do a little bit of research up front. Let's not just dive into building, although that would be very exciting. so, why don't we ask look at the speechtoech benchmarks and find out who are the top model providers for speech to speech.

22:21 So we want to be able to figure out sort of which model how we want to actually implement that. Let's go figure out who is top here. And while we do that maybe we could work with Ashley and ask a few data questions. >> Yeah, sounds good. So I guess let's switch over here. So yeah, I think the things that we I'm thinking of the things that we might want to be that we would want as input when when thinking about a new feature and maybe some things we could start with is understanding the market size of the product and kind of the opportunity. So we could start by asking help me understand the market size and opportunity for voice products. I'm interested in TAM. yeah, let's start with that.

23:04 and one of the things that I really love about having a purpose-built sort of analyst or data focused bot is that, over time just gets really smart about how you like your data. we actually have some custom skills here that help us format data in a really nice way and make it kind of match our match our brand. So when we're going into you know putting together slides for a meeting or kind of be able to share data with your colleagues it kind of looks and feels like that what everyone's used to. I also think every organization and every team you know does data differently right you have different KPIs different metrics different maybe you're you're annualizing your revenue in a different way. And so having a bot that can sort of be that go-to for all that I think is is really great and allows you to kind of establish some of those patterns. So it looks like Ashley is grinding away here.

23:52 >> Yeah. Yeah. Yeah. running some queries for us. the other thing that I love about Crockbot is it can multit these agents can multitask and so you can ask multiple questions in parallel and so while Ashley goes off and does this research for us we can also ask additional things like we have a bunch of voice capabilities already across our product. So, show us what percent of users use any voice mode across our products.

24:24 >> And while we were typing that, Ashley came back with some insights. >> Yeah, looks like we have maybe a 48 billion market ahead of us. That sounds pretty nice. and yeah, 12.5 billion SAM. And it looks like Ashley here because over time we've we've taught Ashley how to make charts that we like has has put together some charts for us. So some some analysis here on the market size and it looks like we're seeing well voice API and infra is pretty close to Devi and coding agents. So nice opportunity for us and obviously yeah I think there's some really good stuff there. yeah looks like we have some some keer data too. So all right seems seems like a worthwhile opportunity I think.

25:00 >> Very cool. And one very useful pro tip is to always ask the agent to bring its citations. And so that way you can inspect and maybe there's some additional editing or follow-up questions that you want to provide on the underlying data set that agent had used. and so oh look at that. Okay. So we got voice tam. Okay. So this is a visualization for that voice tam here. And then we get the percentage of voice weekly active users. Yeah. Across our different products.

25:34 >> And this is the kind of data that I think can be really hard to obtain sometimes by yourself when you're just a solo PM or a solo product person. You're often reaching out to a data team. You're, you know, making sure you can query the right databases in the right ways. And so, I use this kind of stuff all the time. And it's a really great way to say, okay, cool. So it looks like we maybe you know see good penetration of voice products across our our voice use voice features across our product surface. maybe makes sense for us to double down there with the feature. So >> yeah very good. we'll also call out that Grockbot right now internally in SpaceX AAI is sitting on a great foundation. And so one of the things which comes through in this example is you really need a good clean data set >> and a good data lake to to connect Grockbot to. And so there's been a lot of hard work on our data engineering team who are using Grockbot to build a lot of these data sets and curate a lot of this good data on top of which Ashley can query and provide trusted insights. Exactly. There's a lot happening underneath the surface to make sure that you have the right foundations for agents to be most successful for your product team.

26:43 totally like having canonical data tables, making sure that there are some skills made available to the agents to know how to avoid common querying mistakes and all that stuff that you can sort of obviously encode into these agents. But it looks like PMP Pete's come back with some data. Yeah, let's pop back. we were asking who's top of the leaderboard for speech to speech. it looks like Grock Voice ThinkFast 2.0 is top of the Frontier native audio.

27:07 So maybe we should use that model. Sounds good. >> Okay. So I think frontier models are great. So yeah, we should use them. >> PMP is also equipped with a skill and a template for our product requirement docs. And so we have a weekly product review that we can bring any open topic or any new idea that we want to discuss as a product team. So should we work with PMP to to write one of those?

27:32 >> Yeah, let's do it. We should ask ask PMP to help us. Let's write a prd for voice mode in crockpot. And I think he's already connected to notion, right? So this will get put into our notion. >> That's right. Okay. Awesome. Awesome. yeah, so PMP here has context on how to write really great PRDs. And so PMP knows that a great PRD is not just a bunch of requirements on a doc. It's insight from user research which we just looked at together. It's maybe insight from the industry and landscape that maybe Ashley just helped us put together. and obviously thinking about prototypes and increasingly we find ourselves prototyping product a lot early in the process for how a you know how something could look and feel in the product. So we'll see what Pete comes up with. is there something else we should show in the in the meantime here?

28:25 >> why don't we have Pixel start cooking on some design? >> That's a great idea. Yeah, actually I might just ask Pete to do that. So, one of the cool things about Grogbot, can you also loop in Pixel to help with design. so yeah, you can just message message these folks. They'll they'll keep working in the background and can handle the multitasking, but you can also tag other bots. And I find this to be really fun. And so, Pete here, oh, there we go, is going to loop in Pixel on voice mode design, and is still drafting the PRD. So, Pixel, as we talked about earlier, is our designer who's been trained on design systems and how great design works, how to think about iterative design, and design thinking. obviously, a lot of that knowledge has come from an amazing set of human designers to help help make it really great. but in our case, we can really get a head start on some of this process. And you'll see here's Pete's doing >> and we can kind of spy on their conversation a bit. So, it looks like there's some context that shared.

29:27 some seed strings, a visual brief, some product review, and a mock section that looks like it's empty that will hopefully get filled out. So, all right, we can close our readonly chat here. Looks like pixels briefed. >> Amazing. Yeah, I think anybody who spent a lot of time thinking about how best to prompt agents. One of the things you learn is agents are actually very good at prompting other agents and working with other agents. And so we have just put these two into a group chat and PM Pete is able to actually share all of this context and some of the things that PM Pete has learned directly with Pixel without us having to copy and paste or move context around. Agents can just share that behind the scenes.

30:12 >> I think one of the cool things here too is it's they're sharing but also each agent still has its own memories, its own context, its own routines. And so, you know, if we've really worked with Pixel over time to teach its specific behaviors, let's say, you know, always use dark mode, Pixel's going to remember that even if Pete is the one messaging Pixel in this case. And so, I think that also just takes a lot of the mental load off. You can kind of, you know, enforce your preferences once and then have those preferences be reflected on all the future work you do. And it sort of like really makes that initial setup worth it, I think.

30:41 >> Yeah. And are there specific routines that you like to set up across your analyst? >> Yeah, so one of the ones that I really like and while we're cooking on that that doc is I really love having regular product monitoring for new features. And so you know as we're launching new features into Grockbot, being able to see how they're going is really helpful to the product team so we can you know both just keep an eye on usage but also just know where we should double down our investment. And so, one thing that I'll do very often that we can ask Ashley to help out with here is, we recently launched support for the XMCP in Grockbot. We did that on Saturday, actually. can you send me an hourly pulse of how installs look?

31:26 And to start, give me the last few days of usage. and so this is me setting up a routine here. I'll do this very often for new product launches. And so every morning I'll wake up and have maybe well for the hourly ones I'll see them every hour, but also do like a morning brief and kind of see how are all my products doing where like where where am I seeing more usage if I see a metrics dip.

31:50 What's really great about this is I can also just instantly sort of in the chat be able to inspect and analyze what what what happened. what's really great about analytics use cases like this is because it's tied to the data lakeink as Kevin mentioned the data warehouse under the hood. we can also look at product metrics from other parts of our product suite to try and explain what's going on. And so, yeah, it looks like we have a routine that's created.

32:13 That's our hourly pulse. Looks like some installs happening. And there's our Yeah, there's our every hour. >> Very cool. >> Very cool. Yeah. So, yeah, and this is this can be just a really great way, I think, to get analytics to be more of a passive part of your day-to-day. before something like this, I'd find myself just like opening a dashboard all the time, refreshing it, keeping it pinned, having it on a wall somewhere. and now I just use my bots. So very cool. And while we were setting up that routine, PMP has returned with our product review doc.

32:46 >> and so we have a draft of our Grockbot voice mode. Let's take a look at this. so here is the opportunity. you could even pull in those TAM insights from Ashley. Here are some of the customer pain points pulled from social media and the requests. Didn't have access to the chat. Y >> here, but that's okay. and here is the proposed solution here. So live Grock mode, voice mode on Grockbot, which is very cool.

33:17 >> I like this. Yeah. okay, looks like there is still a placeholder for Pixel to pop in some mocks, but it it knows that Pixel needs to do it, which is really nice. So, we could always ask RPM to make sure that that gets completed. and I really like this, too. So, sort of like here's a doc and here's a TLDDR. So, I can refine the TLDDR a little bit. Oh, wow. There we go. There are our mocks. Just I was just saying I was just saying that. Yeah.

33:38 >> So, Pixel has just handed the mocks back to PM Pete. >> Let's see how they did. >> Amazing. Look at this story board. >> Oh, yeah. I like this a little little bot in the bottom. >> Okay. Whoa. Went with the huddle director. Very interesting. >> Yeah. Yeah, I mean I do that a lot in Slack, right? We we we huddle each other. So >> listening little back and forth. >> Okay, look little menu here.

34:04 >> Menu crisp. I think >> Oh, even went to mobile. >> Even went to mobile. >> Started from action button. Okay. Yeah, I mean this was like a really good starting point, right? Probably could use a little bit of polish and consistency, but definitely enough to make a great prototype. Should we try to build it? >> Sure, let's do it. >> so Emily's our engineering manager, right? That's right. All right. And then what do we think? We should ask Pete to ask Emily to build it or just go straight to Emily.

34:28 >> Let's just have them work together. >> Okay. So, one thing that you can do in Grockbot is you can actually just create a group chat. So, we'll add PM Pete and where is our There's Emily. we'll just do our PM and EM band up, right? and say, "Hey guys, hey folks. >> Pete just put together a purity for voice mode. Can y'all sync up and Emily start building this?

35:02 And so yeah, this is now you can see that we're in a we're in a group chat with Pete, Emily, and ourselves here. And what's really cool about these group chats is that they feel just like talking to a bunch of colleagues at once. And so we're getting the contacts from both of these agents. Obviously earlier we saw agents can sort of tag each other and and prompt each other. but we'll I'll often use something like this as a way to just be able to see the context for both agents in one place and be a little more hands-on, right? So, sometimes I'll, you know, ask my chief of staff to you know, talk to a few other agents to get something done. For I think build cases like this, it's really nice to be hands-on and get the get the agents to be talking to each other. So, cool. Very cool.

35:44 And we will let them work together down here. And what Emily is going to do as a next step is actually distribute this work. So we'll take the product requirements doc or RFC translate that into discrete technical tasks which then she will hand off to her team of engineers who will each get a ticket. Let's see. Oh yeah, looks like they're chatting, checking on them. There's Pete confirming the source of truth. And then Pete, who's put together our requirements, has the V1.

36:22 Any good product starts with a V1 and MVP prototype. And so looks like Pete's got that figured out. And they're thinking about it. Yeah. So your stuff that's out of scope, which I love. Always always great to see that. Cool. And they should be cooking. Very good. And one of the amazing things of Grockbot is it is able to actually spin off cloud agents to do coding tasks. And so you can think of Grockbot as being able to hook into a specific coding harness and a set of agent cloud agent environments which is really optimized for coding tasks. And so you can have this great handoff and these managers of cloud agents going forward.

37:07 Yeah, I think this goes back to what we were talking about earlier with what we saw internally, which is a lot of our really great engineers found themselves managing multiple sort of even dozens or hundreds of agent threads over the course of a day and sort of were creating their own versions of orchestrator or coordinator or sort of manager agents. and so one of the ideas that inspired this particular feature in Rockbot was making that feel really native. and you know there's a bunch of debate and discourse around how much people are looking at code these days versus you know looking at outputs and outcomes and I I think this actually feels like it could be a really you know interesting way to write code in the future and and even now but basically you know focusing on outcomes asking agents to coordinate with each other and then sort of giving agents the ability to go and spin off these these you know coding focused environments where they can get a lot of things done so yeah looks like they're cracking way.

38:02 >> Very good. So, it looks like Emily has actually given me a task to unblock one of the engineers. looks like there is an approval that is required on my side. And I think this is a good example where these types of handoffs. agents are really trying to push as to be as active as possible on your behalf. but they can sometimes run into access issues or need your help to log into a tool that you use. And we want to equip our agents with the exact same tool set that we have access to so they can be most productive. And so we'll often get questions from the team here asking me to log in or whether they can have access to a particular tool that would make their work even more productive. Yeah, I think it's really helpful as we talked about earlier to make sure everything's connected.

38:52 Obviously, over the course of writing software, depending on a feature you're writing, might need to be able to valid validate those in different ways. and so looks like our engineers are cracking away here. one thing I'll also mention that we didn't talk about much earlier is you can do all of this on mobile, right? So, we saw that mockup of Grockbot on mobile voice, but I'll often do this on the go where if I have an idea for a feature, I used to, you know, pull up a notes app and then wait till I got back at my desk and then kind of type that into a prompt and then write a doc. but I can kick off this entire engine when I'm just on the go.

39:24 you know, just I can talk into Grockbot with the the speech to text flow and maybe soon a voice flow. but really being able to access this anywhere on and on any device, I think has been a really huge unlock. like sometimes inspiration strikes at the end of a meeting and you just want to be able to kick off an idea or kick off a prototype and once you have the system set up you can really go from zero to prototype really really quickly. So very cool and so it looks like Eileene and Emily have who's >> Eileen Eileen's one of our >> one of our engineers and Larry are all coordinating on this work and so looks like they are looking into the best way of implementing pool hint.

40:05 and so this is just a great example here of really the autonomy and ways in which grabbots can coordinate their work work together on some of these larger features. Yeah. And if you want to set up something like this, I think you know you're seeing that a bunch of different people or agents are being pulled into the chat. And that's because the way that Emily is set up and these engineers are set up is that they know they can tag each other. So, over time, as Emily has grown in in her role as as the EM here, she's learned that she can task out certain things to sub agents or other agents, that she can loop in folks to help delegate work and then be able to review that work. So, sort of creating like a task and reviewer paradigm, and then also be able to sort of sort of test features independently as well. And so, we didn't in this case didn't have to tell Emily to go and tag these other bots, but that sort of happened organically as a result of how we built this. And so, you know, for your development workflows, if you're excited to use Grapot for this, obviously that might look different. That might look like one, that might look like two EM that might look three EM that might look like just a fleet of individual engineers where you're the EM role.

41:10 but really the point here is that you can set these bots up to do whatever makes the most sense for your workflows. and then kind of have that pattern be enforced on on future runs. So, how far do you think we'll get? >> I don't know. We'll have to wait and see where this team ends, but it is very cool to see them in action. I also think one of the things that we have thought a lot about is sort of at what point should you as the coordinator and PM on the team actually take back control? And I really like to think about really offloading a lot of the toil and maybe lower sophistication, lower quality work to my agents and sort of building up context as we go. still being sort of that last mile of reviewing the you know research that agents have put together.

42:00 refining that product dock that we saw to really make sure that it represents like the best strategy and the best product that we could potentially build. And then obviously there is a lot of code review that goes into ensuring that the code that our our agents are writing can actually make it into production and really thinking about that full pipeline from idea through to code in production and what are the sets of systems and agents and bots that you want to bring to bear to really accelerate you.

42:30 >> Yeah, I think that review and refine is a really good way to put that. I think I find myself doing more of that work and I think there's always some stuff I catch and so I feel like it is a really important part step in the process here. But I think what's really changed with with something like this is that I can I can kick off a lot more work than I used to and it feels like I can manage a lot more work than I used to because I can delegate a lot more than I used to. And so that that sort of freeing up more time for me to focus on review actually feels like where I want to spend the most time as a product person. like I want to be thinking about what's next and trying out new ideas and I don't want to be thinking about, you know, the GitOps under the hood for me to be able to make sure I'm on latest for the design system changes that went in last week. Like I want that to be abstracted away. And so, I think one of the really powerful things here is you can do that. You can focus on just being this creative product person, who's thinking about what you want to build next and doing that as quickly as you can. So, >> very cool. Well, I want to make sure that we have time for questions and discussion at the end, but I do see one other agent here, which looks like Xbot. What is Xbot?

43:32 >> What do you think Xbot is? Xbot is actually, a, agent which, manages my, XDM. And so what we are going to do is for the next 10 minutes, anyone who DMs me N2 Parkco on X, let's send them a link which will have a $100 promo in credits. So let's fire that off and let's let Xbot respond to anybody in my DMs with a little promo code for Grockbot.

44:03 >> Awesome. And yeah, these bots can do so much. They have so much ability to connect to services and so yeah, that's it's great example. Very cool. Cool. >> and for anybody who's not streaming on X, you can also scan this QR code for joining us today. would love to hear what you build. and very excited to get Grockot in your hands. >> Yeah, thanks so much. we'll see if we can get a version of what we just built out maybe after the after the call. But I think the takeaway here is that these these systems are really powerful and they're systems that are really u the sky's is the limit. And we've seen all kinds of different teams put together.

44:39 if you're interested in seeing this team in a bit more of a text format, definitely check out Kevin's post on X to take a look at how you could build this. And actually, when Kevin first wrote this, the way that I set up my team to look just like his is I just pointed one of my Grock bots at that post. And so you can you could do that, too. If you want this team to power your workflows, point them at that post.

44:58 They'll set this all up for you, ask you some questions about how you do your work, and you're off to the races. But yeah, very excited to see what you build. Very excited for everything that's to come in Grockpot. And thanks for joining us. >> Amazing. All right, Noah, I believe you have been curating some of the top questions coming in as we went. Would love to hear what's top of mind for the group. There's a lot in here around just what does the future of Grockbot look as far as the platform itself, but also how does that connect with things like cursor cloud agents and stuff like that?

45:34 it'd be great if you guys could talk through maybe how do you see this kind of evolving over time? >> Yeah, I I can start out with that one. really I think we want Grockbot to be the platform for you to do your you know do your most useful work. And so the the brief really is those two things that we showed at the beginning which is focusing on giving you a really useful agenda colleague. but the meta goal here is making it so that you can do more be more productive and you know get more done you know feel feel better about the work you do. and so you know I I think everything we're building is in service of that mission which is really helping you get things done.

46:10 And so obviously for coding workflows that means things like cloud agents and making that experience really great. obviously like a lot of stuff just around how the products interact together that we that we're thinking about. but then I think there's a whole bunch more, right? Like there's the coding loop, but then also the rest of the system. And I think excited to explore more stuff there. But long story short, you know, we talked earlier and showed you earlier how these agents, these bots can create cloud agents.

46:35 We're really excited about that capability and continuing to support it. And so yeah, I don't know if you'd add anything on the cloud agent side. >> Yeah, I think there there are sort of two big dimensions. one is increasing access for bots and so ways in which you can work with bots in Grock bots in new ways. so one example of this is we just introduced our Android app this morning. and so bringing this to new surfaces, new platforms and being able to work with your team just like you do colleagues I think is a a really important tenant. and then the other is allowing our team here to be able to do more. And so a lot of that means new tools, new connectors and new ways of working that are going to allow your Grockbot team to be able to take on more and higher complexity work over time.

47:30 another common question comes around some of the context management piece and I know that that's like very different than how context management has typically been done within cursor and many other platforms. Be helpful if you guys might have touched on this a little bit but to frame some of the mental models people can think about when effectively like managing the context of these agents since it is such a different paradigm. Yeah, we spent a lot of time thinking about call it like the memory system and routine system for these agents. and particularly because they behave, I think differently than a lot of other agent products do. They, you know, have their own computer. They have a persistent computer. They can kind of execute code in that environment, execute tasks in that environment, save files to disk. and so I would say from a context perspective, the vision here is that you should never have to worry about context, right? And I think that's maybe the again like the way we're thinking about this is sort of historically you're always watching a little meter or thinking about how much you're using for a particular chat. And I think the goal here is that these are very sort of incredibly longunning agents. Like my conversation with my chief of staff has been running for months. and so really you shouldn't the goal here is that you shouldn't have to worry about managing that context and that these agents should get smarter over time and at least from our usage internally that longitudinal behavior proves out right I work with my agent who remembers the stuff that's most recent some of the older stuff that's like kind of fallen off the road map list for example like my bot doesn't bug me about that and that's kind of what I want and so really I would say that that's at a at a broad level how we're thinking about it and obviously got a lot more a lot more to there. So yeah, and anybody who's invested heavily in agent context systems in your skills, all of that ports over to Grockbot. the other thing which is relatively unique and something that may require a bit of a adaptation if for those who are used to working with agents exclusively via the textbox is your Grockbot team can learn on the job.

49:32 And so, being able to teach it, correct it, show it things, actually, you can, record, workflow on its computer and it can learn from the ways in which you work. And so, there's sort of this whole new dimension that opens up as, your team becomes more autonomous of ways in which you can teach it and extend it and make it better and better as you do your work. There's a lot of curiosity just into like how our team is using Grockbot as a whole. people have asked like what portion of like cursor or SpaceX in general is using Grockbot in their work.

50:11 Know we talked a lot about the PM use cases here, but there's also questions of like, hey, how much of of the features that get shipped are designed by Grockbot versus you all? it'd be great to maybe pull back the curtain on just like the the general usage there. >> Yeah, I think one of the things that we saw really from this product's inception was internal product market fit which is really a good signal. and so we have you know a a large go-to market team very resonant among different parts of our go to market team from finance to our sales team to our post sales team being able to orchestrate a lot of what goes into account management to our sales pipeline and even to running our our business. And so yeah from the get-go it was high levels of of product market fit. The other thing which I think it was sort of novel here and and not expected is just how useful Grockbot has been for product and engineering teams and technical teams. And so really the connection to cloud agents and having a really great coding harness alongside a autonomous colleague has been really magical.

51:28 And we've seen a significant number of our merge pull requests these days are now coming from colleagues in Grockbot actually initiating the work with cloud agents. >> Yeah, I think that's super true. And then you know Noah to your point around pulling back the curtain a little bit. I think a lot of the ways that we find ourselves interacting with Grockbots is sort of through those connections that we've been talking about. So a lot of times folks will have a Grockbot helping them monitor Slack. And so, you know, no matter what work you're doing, there's like tools that you sort of there's like that the stuff you need to do to just do your work, right? and so there's things like connecting to Slack or, you know, using a particular like HR system or whatever. And so, I think we we see a lot of use cases where folks will monitor the most important issues in Slack and so then know when they need to jump in there and their bot helps with that triage. So, regardless of role, there's there's that there's that type of use case and then you know, I think people are also having a lot of fun with it, right? there's a lot of ways that we're you know adding you know communicating with each other in in Slack and other channels and and I think that's also really great but yeah in general I'd say there's there's a lot of connecting to other system stuff like a lot of our interactions on linear and ticketing are happening with bot like bots initiating those tickets a lot of times like when I'm like looking at the queue of work that I need to burn down in a given week I'm usually using a bot to help me triage that and so I think you know across the across the suite but also for specific tools that's been really cool. So, >> when are multiple emails coming?

52:59 When can people have multiple accounts? >> We're thinking about it. nothing to share or promise yet, but definitely hear the feature requests. and please do keep those coming. Like our team is all over X and all over like the feedback we're seeing in the industry and it it really really is helpful. where we're like what you just saw where we'll look at a how how often a feature is used and get some some usage data on it. it's something we're doing all the time like we are truly doing that hourly and daily internally. and so yeah, please do keep those coming. and we're we we hope we'll have a lot of exciting stuff coming soon. So >> yeah. Yeah. And feel free to reach out to us anytime as you all get going. We'd love to hear your experiences. share the bots that you're building. and would love to hear any feature requests.

53:46 So, yeah, thank you so much for joining us today and excited to see what you build.