GrokBot Course

What is GrokBot? It's basically the way I like to describe it, a team of always-on agents that act at your delegation or your disposal. And what I like about GrokBot is kind of three main things. The first is that each bot has memory.

And so what that means is you basically have a persistent long-living agent at your disposal that you can ask questions to, that will remember kind of the answers it's given before. It will also remember, you know, any follow-ups that you've talked to it about. And I like this because you can have basically a long-running conversation with an agent over weeks, months, and it'll still be able to categorize and understand everything you've talked about with it before. GrokBot also has access to tools.

So skills, plugins, all of the paradigms that you've seen. Yeah, Nate? Are you sharing something? Oh my gosh, I'm so sorry.

No, I'm not. Let me share. Good call, Sankar. Sweet.

Perfect. Thank you.

Here we go. So welcome to GrokBot. It's basically your team of always-on agents, as I've mentioned. And so all of these agents that you have at your disposal are basically going to be agents that have long-living memory, as well as tools and computers at their disposal.

So you'll see that all of these agents will actually live inside of your GrokBot interface that I'll show shortly. And what's nice about these agents is that they can actually communicate and talk to each other. So these three things is what I want you all to take away from this session is that an agent will have, or your bot will have, persistent memory. It'll be integrated in your ecosystem, and it will also have access to its own computer.

And I'll show all of these three things today so you can kind of see what makes GrokBot so special and so different. The main thing that I want to hone in on today, and the main difference with GrokBot, is we were in a world where a lot of times we were just asking agents to do things for us. So we were just saying, hey, can you create these slides? Hey, can you edit this file for me?

And we're going more from a task-based mental model to a persona-based mental model. So a lot of these bots that live inside of your ecosystem, they have personalities, they have roles, they are experts in particular things. It's been really fun seeing how people name their different GrokBots, which has been great. And so that's kind of a good example of how to think about this new world where you have a team of agents that are working on different things.

And then, as I mentioned, you're gonna go more from asking your agent to do things to delegation. And we'll see that in a very kind of easy example where I'm actually gonna bring up a Chief of Staff bot, and that bot is actually gonna delegate to other bots that exist inside of my ecosystem. So again, what I want to mention today is that these bots are meant to be almost like colleagues or teammates, and they're meant to use your tools like you would use them. So for today, we're gonna start, again, with a Chief of Staff bot, and that bot is gonna have three bots that basically almost report to it, but that it works well with.

So the first is an inbox manager, the second is a schedule coordinator, or think about it as like a calendar manager, and then the last one is a to-do list organizer. And what's nice about all three of these, oh, sorry, what's nice about all three of these is that all of them are gonna be able to talk not only to the Chief of Staff, but also to each other. So I'm gonna show this in action, and then I'm gonna pause for questions after the initial demo, and then show you all a little bit about the controls and guardrails that we put around Grokbot as well. Okay, so let me minimize this.

And this is what I'm gonna show you first, which is just, I have like a sample calendar, so I'm gonna open up my calendar here. And you can see, I just have some stuff going on today. So I have commute, I have some workout, I have some customer calls and product syncs. And what I'm gonna do is I'm gonna have my Chief of Staff, Grokbot, actually help me organize my calendar around to-do lists that I have to do and around kind of what's in my inbox right now.

So I'm gonna go minimize this and actually open up Grokbot for you all so you all can see it for the first time. So this is what a lot of you will see when you download Grokbot. It looks very similar to iMessage or maybe like a chat-based interface that you've seen before, but we try and keep it as simple as possible because we basically have thought that you wanna work with these agents like you would with a friend, a colleague. And so we kind of created that interface to be similar.

What I'm gonna do is I'm gonna create a new bot. I'm gonna create my Chief of Staff bot. And this is kind of how a lot of folks really start out with Grokbot is just working and seeing what their paradigm for organization is. So in here, I'm actually going to click on this bot.

I'm gonna call it Chief of Staff and I'm gonna say, you're my day-to-day organizer, helps me stay on top of things. And you can see it immediately responds to what I name it. So it's able to understand, okay, this is my job and it's gonna kind of like point me in the right direction of does it, you know, do I want it to run the day? Do I want it to hurt other agents or something more specific?

What I'm gonna do is I'm actually gonna ask it to spin up those three or two other agents that I liked. So I'm gonna say, I would like you to spin up a inbox manager, a calendar scheduler, and a, what was my last one? A to-do list organizer and work with those bots. And what's nice about this is I can always talk to my agent as it's working.

So I can even say, make sure that you're able to access my calendar in my personal Gmail account. So while it's working, let me actually show you how I connect to some of the tools and plugins that actually exist within Grokbot. So you'll see here that there's just one kind of entry point, it's called plugins. And you'll notice that I have a couple that's connected already.

So I have Gmail, I've Google Drive, I've Calendar, other common ones that people have are Notion, Slack. But depending on your domain and your role, you can also connect to Salesforce, to Clay, to HubSpot. If you're an engineer, obviously you can connect to GitHub, GitLab, all of these other tools that exist. And what's nice about this is that it is such an easy way to connect that you can also even connect multiple accounts.

So I can have my personal email, maybe my shared calendar with my partner, my work email, all connected in the same Grokbot instance. And it would be able to identify between the different accounts. Awesome, so what you'll see here, and this is so cool, is that it has my three different bots here.

They're all different colors, different shapes. And then my chief of staff actually has a little staff group chat with all of them. So I can see here that all of them are kind of, can talk to each other, can work on things together. And so it's kind of up to me if I wanna put stuff in this group chat here, but I can also have my chief of staff be the one doing the delegating.

So you can see here that it's connected to my personal email plus my family calendar and my calendar with my partner. And so it's gonna ask me to maybe work on a particular task related to my calendar and to my to-do list. So what I'm gonna do is I'm gonna keep talking to my chief of staff here. And I'm gonna say, I'm actually gonna use voice mode here.

I'm gonna say, allow, I'm gonna say, I have a ton of to-do lists or a ton of things on my to-do list. I need to do my laundry, I need to call my grandmother. I have to go buy groceries for a big dinner party this weekend. And then tomorrow I am meeting up with a friend to help her with a presentation for work.

So I'm gonna pause, I'm gonna send that over. It's gonna transcribe. So I'm gonna say, I have a ton of things in my to-do list, just like that, take that part out. Oops, sorry, ton of things.

And I'm gonna ask it to work with my calendar scheduler and my to-do list organizer to slot these tasks in when I have time. And notice that you can tag your agents just like you would group chat. I like that because it kind of tells the chief of staff who I want it to coordinate with. A lot of times it'll do that on its own as well, but this is kind of a good way for kind of the chief of staff to realize these are the people that I want you talking to.

And so what we're gonna see is that it's going to actually put all of these things on my calendar. And what I can do if I want to maybe customize this or make it more my own is I can actually teach it the types of things that I like to do at particular times. So for example, I can say, create a skill that shows or that makes sure that my chief of staff always knows that I like to do laundry late at night. So you can see here, I'm asking it to kind of learn from my, what this means is that not only does this agent have memory, but it's able to learn from my preferences.

So this skill, which tends to be an AI paradigm of basically a superpower that the agent has now lives in my chief of staff forever. And it will be able to learn going forward if I ever want to do laundry that it should schedule that late at night. So you can see here that it kind of makes the plan for me. I'm gonna go ahead and see what it messaged my two different bots.

So I can see that it told my calendar scheduler, like, hey, Amrita asked us to slot these tasks. These are my proposed holds. I can also go back and see, you know, what it messaged maybe my to-do organizer. It said that I captured these four tasks and I can actually go and see what those four tasks are.

So this is a very cool kind of new paradigm where we're able to see how agents can delegate to each other, how they can talk to each other. I've even seen agents push back on each other and we'll actually see an example of that in an engineering use case where you have a front end and back end engineer talking to each other. But now if we go to my calendar, let's see how it did. Cool, all right.

So it looks like it did exactly what I want, where it said, I need to call my grandmother here. I think it put my laundry later at night, there it is. And then it said, I need to help my friend with my work presentation and the grocery run for the weekend dinner party is on Saturday. And I like this because, you know, I can change this around, I can customize it even more.

I can make things recurring and I can do all of that through my rock bot. Awesome, I'm gonna pause here.

Nate and Fussell, are there any questions I can answer around this just general delegation framework? I don't think anything specific to delegation framework, but we have just been getting a lot of questions on theme of skills, markdown files, like how that differs from, you know, using the cursor desktop app. And I think you're probably gonna get into it with like teaching grok bot a skill from a video demonstration and sort of how that all works. So I think we'll answer most of those throughout the presentation.

Absolutely, and I'll actually cover skills very quickly here. So we see that I added that skill to my chief of staff, which is the fact that I do laundry late at night. You can see here if I actually click into the skill, skills are very straightforward in grok bot. It basically has a name, a description and then instructions for the agent.

What's nice about these skills is first of all, they're shareable. So it's very easy to actually publish the skill and then maybe give it to Nate or someone else on my team. And then they can use that same skill very easily. Similarly, if you have focused, for example, if I actually show you my full list of skills, go to plugins down here, you can see that I have a skill for my QA agent.

She's called QA Quincy. And I basically tell her how I want her to do screen recording for a QA test. And you can see here that, you know, I have instructions on what to record, how to do it. But this is an example of, you know, a skill that you could maybe take from cursor or any of your previous AI tools and just plop right into RockBot.

And what Nate was mentioning, which is kind of where we're gonna start talking about computer use, is that you can teach your agent a skill as well. So for example, if I wanna teach my chief of staff agent, let's say how to book a flight for a family vacation I'm going on, I can actually open up my chief of staff's computer. So this is all kind of on a remote VM. Each of my agents, each of my bots here have access to their own computers.

And I can often see it, you know, do things on its own. But what I like to do sometimes is actually teach them how to do something. So for example, I can hit this teach a task button and I can actually click into the browser here and I can say, I want to look at flights from SFO to MIC. And I can basically teach it the types of flights that I like booking.

So for example, I always wanna use United. I always want to fly preferably early in the morning. And then, you know, if there's anything related to the flights itself, the type of seats I want, it's able to kind of understand that. And if I just pause this now and go back here, my chief of staff is now able to learn from this demonstration.

It's able to see how someone can book a flight or how I can book a flight, the type of preferences I have, and it will turn that demo into a skill. So now my chief of staff can, you know, put this flight into my calendar and basically keep learning from all these different workflows that I'm showing it. Awesome.

So a little thing about computer use. Right now, Rockbot has access to all of these plugins, but there are times where there is a plugin that doesn't exist or there's certain things that it can't access via an MCP. That's where computer use really comes into handy. You'll notice that, you know, for if you ever wanna create slides and maybe, I actually don't know if PowerPoint is a plugin here.

Looks like it's not, but you can easily teach your chief of staff how to make PowerPoint decks by just opening up PowerPoint on your Linux VM. You can also teach it how to make slides in Google Slides if you want. And so doing all of those showcasing, doing all of the teaching of skills really makes it easy for Rockbot to be able to use its own computer and do all of these things for you. The last thing that I wanna highlight here, and as I mentioned, Rockbot is available on mobile and it's available on desktop, is that Rockbot operates completely remotely, meaning that I can close my laptop and my chief of staff, my organizer, my calendar scheduler will all still run with no problem.

I don't need to keep my laptop open. I can have all of these agents running in the background. They can run while I'm in a meeting. They can run while I'm away from my keyboard.

And then I can come back and see the output. So what we really see from Rockbot in terms of value and ROI is the fact that you save time. You save so much time just by having these agents work in the background for you and complete things that can be tedious, that can be time-consuming, and that can oftentimes be impossible to fit in a 24-hour day. These agents are able to do all of those things for you.

Okay, one thing that I wanted to show besides skills, so you can see here that it's turning this skill into a flight skill, and we'll come back to this shortly, is a routine.

So you can see here that I have this very simple button just called Create Routine, where I can actually give my Rockbot basically just a recurring or a scheduled set of tasks that I want it to do. And here's the Google Flights skill that it just created from that video I showed it. Amazing.

A very common routine, or one that I actually depend on, is having a routine for my inbox. So I'm gonna go to my inbox manager now. We haven't worked too closely with him yet, but let's see how he does. And what I'm gonna do with him is I'm actually gonna say every morning, so I'm gonna say, I want you to be like my morning briefing routine, I'm gonna say every morning, I want you to go through my inbox.

And I'm someone that subscribes to a lot of different magazines, newspapers, journalistic, you know, Substacks, all of those things, but I always don't have time to check all of them because they're kind of scattered through my inbox in between a bunch of spam. So I'm gonna say, every morning I want you to go through my inbox. I want you to check all the newsletters. I subscribe to, and then I want you to send me a Slack message with the highlights from each newsletter.

And, you know, it can be like Pragmatic Engineer, which I love, it can be the Skim, Current Affairs. There's a lot of different ones that I subscribe to, and so I'm just gonna ask it to basically consolidate all of that information into one place for me, as opposed to me kind of sifting through all of the unreads that I have. You can see here that I have a bunch of different triggers. So this one I would want probably every day, maybe at, you know, 9 a.m., maybe a little earlier, 7.30, but my triggers don't have to be time-based.

They can be based on a message being sent. They can be based on the incident being launched, and I'll actually show an example of that shortly. But cool, I'm gonna make this so that it's active. And then what I'll do actually is I'll just authorize this so you all can see that I'm actually logging in to Gmail for it, so it's able to know that I am giving it permission to access my email.

I'm just gonna go in here, and then that way, it will have access to my inbox, and it's able to see all of these things. And this is exactly what a lot of our sales team does whenever they want to send follow-ups. It's what our marketing team does when they need to send, like, mass emails, is they have kind of RockBot give access to its email. Beautiful.

And then that they're able to run these routines or run these kind of scheduled delegation tasks. I'm gonna just do a test run right now so we can see how this does. While this is running, are there any questions I can answer from the crowd? Yeah, I think a lot of things that I'm seeing now too is related to, like, how do you think about tweaking or changing the memory of your RockBot, which I think you're probably about to maybe get into here with some of the instructions and creating these routines, but that's kind of another theme that I've seen come up in the chat.

Yeah, I do think, so the interesting thing about memory in RockBot is, this is basically the first of its kind in terms of these long-lived agents that have persistent memory. Memory is stored in S3 buckets, so it makes it really, really powerful if you ever need to ask the agent to think about something that it worked on maybe a week or two weeks ago. If you ever want to tweak your agent, I would say that routines, skills, all of the things that you would do normally in Cursor do help kind of make your agent become more powerful. There's also, I think, good examples of how you can tell your agent what not to do.

LLMs, agents generally respond really well to patterns and anti-patterns. So a lot of times when I create skills and there's a particular thing, for example, I never want my agent to, you know, format my emails in a really, like, haphazard way. I can give an example in my skill of what a good email is versus what a bad email is. And so that's kind of a good example of how you can teach your agents to get better and to learn from what they're doing, but they will do that automatically.

I think that's kind of the beauty of GrokBot. One thing I do want to call out though, while this is running, is there's also some controls that you can set with GrokBot in terms of what it can or cannot do just from like a system level. So if I go to my action settings here, you can see, I'm actually going to close this guy. You can see that, you know, I have some general controls on.

So for example, if I don't want GrokBot to, you know, run everything without me telling it to, I can turn this off so that it will check each action before it runs. But I think this is what's really cool, which is I can give it specific instructions on when and how it should ask me for a particular task. So for example, if I want GrokBot never to send an email without my permission, I can say when GrokBot wants to send emails to external vendors, it should always ask first. I can add that as a rule.

I can say when GrokBot wants to add a new calendar event to my shared calendar, not my personal, I can say that it should always ask first. And so you can be, you know, systematic and granular here in terms of, hey, I don't want GrokBot ever to do X, Y, and Z. And so I'm going to add up those rules here. You can see I have some other rules that I'm saying allow automatically. And actually when you use GrokBot for the first time, it will start prompting you being like, hey, do you want me to do this?

Do you want me to always do this? Or do you want me to never do this? And so it's going to go through all of these kind of test cases with you. And then as you kind of use GrokBot more and more, you can start setting those guardrails in the settings automatically.

So GrokBot has memory. GrokBot can learn from its prior mistakes or from what you tell it. And then GrokBot also has controls. So you can make it very clear when it should or should not do something based on what permission you gave it.

Okay, this morning routine is still running.

You can see that I have a lot of unread emails. So it's giving me kind of just like a general, like nothing really needs a reply, a lot of things. So this is just something, a newsletter that I have to reply to between my friend group that I need to do. But other than that, what it's going to do is that it's going to send me that newsletter highlight to Slack based on what I told it here.

And so that's kind of a very cool way for a lot of people to use GrokBot is just as like an email triage or as an email cleaner. A lot of times I've seen people use it as a way to kind of filter the noise and just really take what's important. I've seen people file expenses using the computer use and the routines via GrokBot and not having to do that manually. And so these are really great ways that you can kind of take those minutes off your day that would take so much time and actually just have an agent do a number of them.

Nate and Fuzzle, is there anything that you guys would like to add based on how you guys use GrokBot before I move into the engineering use case here? I think you kind of covered the way I use the inbox and calendar scheduler a lot. So I have my chief of staff and I have these, what I call like utility agents. And so I'll have the chief of staff talk to the utility agent so that they can get those tasks done.

And then kind of offhanding that from one agent to another so that they can kind of focus on that for me. And I think that's the theme of how bots talk to each other in GrokBot. Yeah, and in this group chat, which I love, sometimes I like to do is I like to maybe just give it a little bit of leeway and say something like my inbox and my calendar are really noisy. Can you come up with some strategies on how I can better organize my day?

And sometimes I do this when I'm just so overwhelmed with the number of emails I've gotten with the number of calendar events I have, with things that I know have been piling up in terms of, you know, I need to pay Xfinity, I need to respond to X, Y, and Z. This kind of lays things out for me in a really systematic way. So sometimes if you don't even know where to start, so a lot of times if you're, you know, maybe just opening up GrokBot and you're like, I wanna organize my life, how do I do it? Talk to an agent about that. Get the strategy from your agent and then you can start kind of doing the delegation and the organization of your team.

Think about it like these guys are here to help you. They're here to kind of support you. They have access to your tools so that you don't have to go in and do all this for you. I have people that have used GrokBot to help them organize their finances.

And we're actually gonna have a series of workshops over the next coming weeks where you'll see how GrokBot is used in every part of the SpaceX AI org. So how our sales team uses it, how our marketing team uses it, how our FinOps teams uses it. And so this is kind of a cool way for you to see a general purpose, but then also kind of dive deep into, okay, what actually from this can I take into my day-to-day? Cool.

I don't really have anything else to add, but I feel like when I started using GrokBot, I always found it kind of funny to look at the interactions between your GrokBots and like watch their little conversations. So that was something that I know when I started using it, I would check out. And it's kind of funny to see they can get quite testy with each other sometimes. Yeah, they really do have their own personalities.

And it's funny seeing what people name them. You'll see my engineering team is, they have cute little names. But you can see here, this is an example of how a chief of staff will kind of tag the people important. In the previous scenario when I showed, I tagged them myself, but this is a chief of staff tagging them and basically getting responses.

The inbox manager is saying, okay, I'd run like a filter. She sees only mail that needs a reply. This morning there was 200 on red and a lot of them were from LinkedIn. The calendar scheduler is kind of giving her take on this.

So this is kind of a cool way to see what the agents think about what my day-to-day is. And this is where you would input skills here and kind of give it more info about how I like to run my day. Okay, cool.

I'm going to switch gears now and go and talk about how I would extend this to an engineering team. And we can come back and check on these guys at the end. I would actually love to see, oh, it looks like my newsletter highlights are in Slack already. So we can come back and see how that did.

But I do wanna talk a little bit about like, how would you use this at an enterprise or an industry and what it looks like for particular teams. So I'm actually going to close this and I'm going to open up my engineering team. So the cool thing about Grokbot, as I mentioned, is that you can mold it and basically play with it so that it works inside of your guys' workflows. And so this was actually based on how a lot of engineers at Cursor use Grokbot.

And so what I'm gonna show is basically, I have, I think, five teammates here. My main one is tech lead, Taylor. So I'm actually gonna pin her. So she'll be at the top right here.

And she's the one that basically works, again, with her engineering team in order to build a new feature or to basically push out something to production. So all of these engineering personas, what's nice about this is they all have access to the repository I'm working in. It's basically a flight booking repository called Flylo. But what's nice about this is that, again, these are not task-based agents.

These are agents that are personas, meaning that back-end Bobby doesn't just have to work in this one flight booking repo. He's a back-end engineer, meaning that I can give him another repository. I can give him an open source repository. I can give him a repository that he's never seen before, but because he's a back-end engineer, he's gonna be able to learn it, understand it, and work on it.

And so that's kind of a nice thing about these personas is they're able to work across multiple projects, multiple long-lived projects, but still have their memory, their persona, and not have to get that context switching back and forth when you're asking one agent multiple things. So what I'm gonna do is two things with my engineering team. I'm gonna show you all how I would build a new feature. I'm gonna start actually with Tech Lead Taylor on that.

And then I'm gonna show you all what an on-call agent actually looks like. So on-call in engineering is basically people who have to immediately fix a incident or a bug that comes up, and it could be at any time of the day. So I was a software engineer before this role, and there were times where I would have to wake up at 3 a.m. or 4 a.m. to fix an issue that was pulling down on all, or that was basically breaking the connection. And so the nice thing about Rockbot is you don't have to do that.

You can just have an agent that's always on-call, and that agent can be woken up, can be triggered at different points of the day, and be able to kind of solve those problems without kind of disrupting anyone on your engineering team and their workflow. So I'll start first by talking to Tech Lead Taylor. You can see that I actually have a skill for Tech Lead Taylor, which I call the new feature girl. And this basically is a way for me to work with the tech lead on building a new feature and having her ask me questions that maybe I didn't consider.

So sometimes I'll call this a doubles advocate, or I'll call this grill me, which is a very common skill name. And you could basically use it as a thought partner when thinking of a new feature that you want to make. But for today, I turned it off just because she can ask a lot of questions, which is good, I like that. This is me and her going back and forth on some of them actually.

But for today, what I'm gonna do is I'm just gonna ask her to work with a team. So I'm gonna say, work with our engineering team. To build a new feature for the Fly Low, that's our booking website for the Fly Low project. I want to build a way to not just book flights, but also lodging and hotels based on where I'm going.

So this is a pretty big feature. I'm probably gonna need to spin up new endpoints. I'm definitely gonna have brand new UI and new routes and pages I'm gonna have to build. I'm gonna want QA Quincy to definitely test this end to end.

And if I actually click into her here, you can see that she takes videos of everything that's going on. So not just the front end, but she'll also confirm that the backend is working as well by taking screenshots of the network inspector. And so this is a very prime example of why you would want your rock bot to have access to its own computer. You can see that it can launch things on its own local host port.

It's able to take full end to end video. So this is an example of QA Quincy taking a video of the front end and the backend engineer, their PRs working together. And that's kind of another thing I wanna highlight with these engineering teams is that a lot of times, when you're starting on a project, the tech lead kind of works with the agents on who does what, but sometimes there's a bit of a waiting game, right? Where Bobby has to put up APIs, he has to set the contract for the APIs, and then Faye has to wait on what Bobby is doing and then build a UI based on that.

Quincy in this scenario will have to wait to see the backend and the front end finish. But what you can see here is that the tech lead and the QA and backend engineers are able to message each other and kind of see, okay, who is still waiting on what? So the orchestration here and the coordination between the agents is quite seamless. It kind of takes away all of the dead meeting time that you would need.

And you basically have these agents talking to each other, basically being like, hey, have you finished this part yet? No? Okay, cool.

Then I'll work on this instead. You'll also notice I actually had two other agents that are no longer with us, Product Page and Release Riley. So you can see backend Bobby was able to be like, hey, Page and Riley are not here anymore, so I'm just gonna loop in Taylor and Faye. And so that's kind of a way for the agents, again, to stay sharp and be aware of what's going on around them.

They don't just do exactly what's been asked of them, but they're able to coordinate with the other agents as well. So here is Faye. She's messaged Bobby and looks like, you know, they were saying, let's actually go here. Oh, perfect.

Faye said, hey, Taylor assigned lodging as this new thing. Please post the contract when you have it, all the paths, methods, until then, you know, I need to wait. Bobby's starting to work on it. Faye said, okay, got it.

It's gonna be owner only. We're gonna just do the destination lodging for now. And then Bobby said, okay, this is what I'm working on. These are the end points, just so you have an idea.

And now Faye immediately can start on the front end. So it's so quick and easy for these agents to talk to each other, which is really great because you want these different agents to be able to listen and communicate and know when something could be blocked. Okay, any questions, Nate and Fussell, from the audience here?

We got a good one from Saad who said, so technically, if I set this up properly once, I never have to open the Cursor desktop app again. Is my understanding correct? I will say a lot of our engineers at Cursor, Frances Tai, she's a wonderful engineer. She will probably speak on this in a couple of weeks.

She uses Grokbot pretty heavily. She still uses Cursor. A lot of times she'll use Cursor to review code. She'll use Cursor for certain features like Cursor Canvas and some of the other kind of tools that we have.

But you absolutely can use Cursor, can use Cloud Agents, all of the tools that you get within Cursor, you can use from Grokbot, definitely. I will say it's a little bit more about orchestrating your agents rather than the implementation. Like if you look at the code, so both Faye and Bobby will put up PRs and those PRs you can review within Cursor, within GitHub, up to you. So it's a little bit more about what the agents are doing and less so about how they're implementing it and what code they're writing.

That I would still reserve for the Cursor side of things. And actually, speaking of the devil, you can even see here how the UI is being built. So if I hit Open in Cursor, you can actually see the folder that it's working in. I can actually see what the agent is doing, how it's implementing it.

I'll start to see all of the files here on the right. So I'll see all of the diff come up, but you get all of the information from Bobby, from Faye into these Cursor Cloud Agents, which is where the implementation happens. And one other thing that I've seen come up in the chat quite a bit is just asking for more dedicated sessions on this. We're gonna do a few of those, but I did see a few come up around prospecting, go-to-market sales, stuff like that.

And I actually just put a X-post from one of our teammates, Krista, who wrote up a really, really nice sort of download of how she has used Grokbot and how it's kind of transformed how she works at Cursor on our go-to-market team. So if anybody is curious about how that works prior to the session that we'll do later, just check out the link that I just put in the Zoom chat. And she will also be leading a workshop next Wednesday, actually. It's not on the calendar just yet, but Krista will be talking about all of the different bots that she uses for prospecting.

And actually, I'll have a slide at the end of this session where I kind of talk about that. Prospecting, outbounding, doing follow-ups, all of those tools, she's built bots for them. She also has a chief of staff named after her very cute dog. And so you can kind of see those workflows next Wednesday.

And then we're gonna have a series, as I mentioned, of Grokbot workshops in particular. So we'll do one for marketing, we'll do one for finance and RevOps. We'll do one for, I think, like SDRs and just general like BDR-type workflows that can really see the ROI of using these agents. So we'll definitely do a lot more of these like Grokbot for product, Grokbot for X workshops.

Okay, while this is working, I'm gonna work with Taylor on something else.

So this is another cool thing about Grokbot is, Taylor's obviously communicating with these agents, but that's not her full-time job. She can handle other things. And so what I'm gonna talk to her about actually, while she's working with these other agents, is spinning up an on-call agent. So I'm gonna say, hey, Taylor, I want to have an agent that's just dedicated for on-call work.

I want them to be able to monitor Datadog and pager duty constantly, and to let me know if there are any spikes in API errors or customer issues coming up. Let's spin one up. And so this is kind of where, again, you can take multiple entry points here. You can talk to your tech lead on how to build it.

You can actually just go in and build it yourself with create new bot. I like working with a tech lead, because again, with that grill me devil's advocate skill, she asks really good questions back. So it makes me kind of think a little deeper about the type of agent that I want here. But she said, okay, I'll create a dedicated on-call agent pointed at Datadog and pager duty.

As I mentioned, these are MCPs. So you can see that I have the Datadog one added already. And I don't think I have pager duty just yet. Yeah, I would have to add that.

Oh, it looks like it's part of my team, so maybe. But these are, again, tools that you can import in via the plugins here and just add via OAuth. If there's any custom tools that you guys use, for example, incident.io, you can always just ask Cursor or ask RockBot to set that up for you as an MCP, and it will automatically be able to identify that as well. So that's actually another good reason why you may still want to use Cursor here and there is to set up these integrations between your code and MCPs that don't exist here.

The goal is to build these plugins more and more. So you can see here that we have Microsoft 365, but we'll also have more of these tools that make it really easy to use. But at the end of the day, what I do really like is to start with the plugins that exist and then kind of go from there to see what gaps. All right, on-call Ollie is up.

That's so funny that it kept the alliterative names. I think that's again a testament to Grokoff's intelligence. Let me see where on-call Ollie is. So let me actually just message them and say, hey, I'm on-call Ollie.

I'm gonna watch for Fly Low, Datadog and PagerDuty nights and weekends. I'll only ping you if API errors spike or a customer basing incident is actually firing. Amazing.

So this basically, in terms of an ROI here, this basically takes so many hours and lack of sleep off of my engineer's time. Now that I have this on-call agent, it's able to constantly monitor these dashboards and tell me if there's anything that I should be worried about. And this is an example again of Grok by asking permission. So you can see here that says, hey, I wanna create a new standing automation that will fire on PagerDuty events and delegate to another agent.

It's inventing its own routine. So it's asking me, is that okay? I'm gonna say, sure, I'll say allow here. And then this one is to pull Datadog every 15 minutes and message Amrita on findings.

So I'm gonna say allow once here as well. And I can always, again, tweak this. So if I don't want it to pull these services, what I can do is I can say like, hey, I only want you to message me when there's like a specific incident that's fired. So like if a Datadog webhook is triggered for some reason, that's when I want you to message me.

But again, I can work with this on-call agent to see what makes sense here for my purpose. Okay, let's see how we're doing with our lodging UI.

So you can see here that it's building it in this cloud agent. Backend Bobby looks like it finished already or it's basically done so that front-end Fay can actually work on it. QA Quincy is still holding until she gets the PRs from Bobby and Fay. But while this is working, I kinda wanna show you all a previous run of this where we actually implemented a feature together, all of us, on using your credit card points to book flights.

So you can see here that this is Quincy actually, again, recording end-to-end flow of it working. But what Quincy does is it's able to get the PRs from Fay and Bobby. So you can see here immediately Fay is able to get that, you know, front-end quick upgrade. You can see that Bobby is able to send that to Fay.

And so the back and forth between the agents for a fully fleshed out feature. So actually let's go ahead and just open this PR so y'all can see. The back and forth is really great because at the end I can see, okay, this is the QA follow-up. The QA, I think agent also added a comment with the video.

Let me see if it's still there. Maybe not. But that's kind of a good way, again, for your agent to learn what it should be, should and should not be doing with your PR. So for example, when QA Quincy first started, she wasn't taking any screenshots of the network inspector.

And so I asked her to do that and then to add those videos to the PR. So if we actually scroll down all the way here, you can see, okay, the recording is on Fay's PR and the comment is with this particular network clip. So if I actually go here and go back to Fay's PR and I scroll down, here are all the new backend endpoints that were set up by Bobby. And I keep scrolling, keep going.

Here we go. This is my QA screen recording immediately added as a comment from me by QA Quincy. And you can see here that I think it even adds like a little comment at the eight minute mark when the backend network inspector is shown, saying like, hey, I'm confirming that this actually works and this is going to be done as expected. So again, you get a full end-to-end workflow here.

You get this product spec that you're talking about with tech lead Taylor, and then you get your fully fleshed out PR with not just unit tests by backend, but end-to-end tests by your QA Quincy agent as well. And so this is how we love to build out our engineering work. Some other examples of agents you can build besides on-call Ollie, you can have an agent that's like a CI watcher. So Frances, the engineer that I worked closely with, she was saying that she has just like a CI watcher agent that basically checks to see if all the tests and the PR are green.

And then if not, it will keep going in and trying to auto-fix things, and that's kind of a nice way of basically separating your DevOps agent from your backend agent. But there's a bunch of different directions that you can go here. For example, you can set up a product agent that's able to take all of these PRs and turn them into release notes. So I'll actually do that with you all.

Create a new bot. I'm going to call it, I think the previous one I had was called product page, which I like, product page. And I can say something like, takes the new features built by the engineering team and turns them into customer facing release notes. And I can even say the credit card point feature for FlyLo is officially complete.

Can you talk to Bobby? Bobby and Faye. Sorry, Fred and Faye. And get details that we can share with customers.

And when we released both Grokbot and Grok 4.6, we had been using this tool for a while internally. And so I know for a fact that our marketing teams worked very closely with our product teams, which worked very closely with our engineering teams to get like our blog posts, our Twitter posts, everything that we basically had to do that was customer facing, but that needed to reflect some level of technicality about the product we were able to do via Grokbot by these agents all talking to each other. So I actually want to keep this one up so we can see how it does and what it's doing here. Oh, sorry.

But while I'm doing that, I'm gonna basically show one last thing in the three minutes that we have before I'll break for questions. But if there's anything else that folks would like to see about the product, I'm happy to show it. I think a theme was how bots can kind of talk to each other and work together. So is there something you've experienced where you think maybe five or six bots talking to each other maybe gets too complicated?

Is there some kind of benchmark that you recommend of like don't have 12 bots? Yeah, that's a really good question. A lot of times I ask people to try it out and see for themselves. I think me personally, this is probably the limit of the number of agents I want all talking to each other.

And a lot of times what is nice about RockBot is they don't always all have to talk to each other for some tasks to get done. That's kind of where the intelligence of this agent to agent communication is, where if I ask Taylor to be like, hey, I want to build out just the UI for this feature, it's gonna be smart enough to know to talk just to Faye and not to Bobby. And so the agents don't always have to talk to each other. If they do, I do think there is kind of an upper limit on the number of agents that you can have where it does get kind of confusing, but it's up to you.

And again, it's how you want to organize your team. Every org, every industry and every enterprise has their teams organized differently and they have their leadership and their managers organized differently. So what's nice about this and I think what makes RockBot so fun is you are in control. You get to be the CEO here and you get to kind of organize your agents how you like and what makes sense to you.

So a bit of a non-answer there because it's kind of dependent on what you see, but you will notice that the agents are smart enough to know who should pull in at the right time without pulling in everybody. Yeah, and also just to add though, I also have like around 12 to 15 agents as my, you know. Yeah, that fossil is stronger than me. I don't think I can handle that many agents, but yeah.

Cool, so this is an example of a page basically getting the customer copy about the one-time credit card payment feature that we worked on earlier.

What's nice is Bobby was like, hey, by the way, we also just implemented this new feature on destination lodging, don't forget that. And so giving that customer facing piece to product page as well. You can see that QA Quincy is now starting on the QA for backend and front end for the lodging piece, which is great. So it's gonna run the full stack cases with network and use the desktop here.

So let's see actually if I can see it working, not yet. I'll keep this open so we can see what it does. There we go, it's launching those. And then front end Faye is actually asking me if I can push code changes to a PR that other team members can see, which is nice.

This is a good example of Grokbot being like, hey, I wanna make sure you're okay with this. And I'm gonna say, sure, that's totally fine. And so this is kind of where you, you know, talk to your agents and make sure that you are giving them the level of access and permissions that you're comfortable with. Okay, the last thing that I do wanna show is the mobile app.

So again, it's available on the App Store and we highly recommend that folks actually just download it and see for themselves kind of the difference. It's very seamless between the desktop and the mobile app, but obviously you have a little less real estate with the mobile app, just cause it's on your phone. What I'm gonna do actually is I'm going to mirror my phone so you all can see what the mobile app looks like. But again, as I remember, or as I recommend, I highly recommend you all downloading it.

So I'm just gonna connect right here. Give me one second. And here we go. So here are all my agents.

We have Fred and Faye, Quincy, Bobby, and then we also have our product page, On Call Ollie. I can scroll down and see the inbox manager, chief of staff, all the guys I had earlier. But if I click into, for example, Fred and Faye, you can see that it looks really, you know, very, very similar. I can see the messaging between the bots still.

So I can go and see all of that communication end to end. As I mentioned, you can add files or images. So for example, if you want Fred and Faye to copy a particular website or a particular brand style, you can always just add that here. And then voice mode.

This I think is what makes using Grokbot on the go so easy is that you can just talk to it. You can send it a message and then it will immediately start working. So especially for your chief of staff, if I immediately was like, hey, I forgot to buy groceries. Can you please add this to my to-do list?

It will automatically do that. And then we're good to go. So again, really easy to coordinate across all of these agents. Saves you so much time, you know, both from a work standpoint and a personal standpoint.

But then also it's just, it's fun. I think that's what I love most about Grokbot is that it's a really, really fun tool to use. It makes working with agents easy. It makes it interesting.

And also I think it's just a real, I think what I'm happy about is that this is like a really, really big leap in technology and I'm super excited that SpaceX AI and Cursor get to be a part of that. Cool, so this is a mobile app.

Again, you know, pretty straightforward, pretty easy. Launching new bots, new group chats, pretty straightforward. Everything that I showed you earlier is available here in the mobile app. It's just, you know, a little bit constrained vertically.

Okay, before we finish, I do wanna just top off with two last slides as a little leave behinds.

Sorry, let me open my keynote. So what I wanna show is the other use cases. And again, as I mentioned, we're gonna have specific Grokbot workshops on this. So in the marketing use case, these are two examples where you can have a media manager and a growth manager that, again, works across your ecosystem.

For sales, you can have a prospecting bot and an outbound bot. And these are, again, just the beginning. There's so many other use cases that you can think of, I'm sure, for folks who live in these domains. And then the last thing is, what did we cover today in terms of what makes Grokbot different and so powerful?

I think it's, again, the fact that bots that act like teammates. These are colleagues that can talk to you. They can talk to each other. They can talk to their superior.

The persistent memory and continuously learning. These are bots that are able to remember what they did before. For example, when backend Bobby was talking to product pay, hey, by the way, we worked on this new thing. Did you forget about that?

That's something that we should keep in mind. And then lastly, the fact that work is constantly happening, even when the laptop is closed. So you can launch Grokbot from your phone. You can launch it from your laptop, but then you can go ahead and close your laptop, be done for the day, and these agents will still continue on.

So a very cool product. Very excited to share this. And I'm excited to see how you guys are gonna use it. As I mentioned, if you want to use it, you would want to go to x.ai slash bot.

You'd either download it for Mac or on your phone. And these are kind of the subscriptions that you will need in order to try it out. But please give us feedback. It is a really, really cool feature.

And we've seen a ton of different use cases on Twitter. There's a one that I love about someone who is now running their entire plumbing company via Grokbot and has saved them hours of their day. So there's really a lot to uncover here. Okay, I'm gonna stop sharing now.

Actually, you know what? I'll keep Grokbot on the screen so that y'all can maybe just continue to see it work. Oops, wrong Grokbot. Let's go back here.

There we go. I'll keep it on the screen, but I'm happy to open it up for questions the last five minutes that we have. Are there any themes, any things that I can help answer? I think we've gotten a number of questions around, well, first, checking out the recording of this.

I'll go ahead and put a link to our workshops page on kirscher.com. So this is where you'll be able to see, Rita mentioned that Krista from our go-to-market team is gonna be doing another workshop next week on how to use Grokbot for go-to-market use cases. So that will all be available on the link I just put into the chat. And this is also where you can find previous workshops that we've held.

So if you guys are curious about checking out any other, like, on-demand content, the recording should all be posted there. And then I believe we should follow up if you input your email into the Luma invite with an email when the recording is available so you guys will get notified. But if you're ever curious about things you maybe didn't intend or missed in the past, a lot of those are available on the link I just put in the chat as well. Amazing.

I'm looking at some questions here. Again, as Nate mentioned, we will send the recording out. People are asking the best way to improve self or have self-improving loops to the Grokbot. The nice thing is that Grokbot will do that for you.

So it will learn from itself and continuously kind of get better as you use it. But adding skills, adding routines, things that will basically make it more customizable to you will probably help. Some other questions are, can Grokbot use your phone to make calls? I do know a number of people that have connected Grokbot to their iMessage API, so they can actually have a scheduled Grokbot that texts their roommates to do the trash or something like that.

And so I'm sure maybe you can probably wire it up to make phone calls, but I think the iMessage use case is a very popular one. And then Sean asked about some sort of token context window limit. There is no context window limit in Grokbot because of the fact that memory is stored in S3 buckets. And so you don't have to worry about context ever getting degraded here.

Okay.

If folks have used Grokbot already, I would love for folks to put in the chat what they have been using it for. If not, I'm happy to... Yeah, I see that someone has a raised hand if they wanna come off mute and ask questions. I'm happy to take any live questions for the last couple minutes that we have.

It could also... Oh, yeah. A couple of live questions. Okay.

Alec, I think I saw your hand first. Do you wanna come off mute here? Hi, Amrita.

I had just a quick question about the continual learning and persistent memory thing. So is it similar to the way that humans work where it's gonna store core concepts and things like that? Like almost like episodic memory instead of having, oh my gosh, I can't think of the word right now, but like picture perfect memory like LLMs have so that it's more so relying on its abilities than it is its memory. And then the second question that kind of as a follow-up to that is whenever you're working on new tasks, how does it...

Is it just that that first time that you do it, is it gonna kind of be rough where you're gonna have to babysit it along the way? And then after that, it's going to continue to get better? Or is it something that it can essentially figure it out and one-shot the task that it hasn't done in the past? Because I know Richard Sutton kind of talked about that on the Sequoia Capital podcast, I think a little while ago.

And so I'm just curious how Grogbot kind of fits into those two aspects of things. Great questions.

I'll answer the second one first, which is I think with most agents, the cold start problem is probably gonna be there regardless where when you talk to an agent for the first time, you might have to do a little bit of hand-holding. You kind of saw it a little bit when the chief of staff agent was like, hey, I'm a chief of staff, but I don't really know what exactly you want me to do here. I do think as, and then that kind of leads into your second question, which is the memory. I do think right now the memory that we have is like a more picture-perfect LLM type memory where you can teach it skills, you can teach it videos, you can give it information that it's able to just consume based on what it's able to see.

I do think the goal, however, is to improve memory as like our number one feature so that it remembers things like a human would. So episodic memory or even like memory related to things that it's seen or listened to. So the concept mapping idea that a lot of humans have. So I think that is the goal is to improve memory so that it's not just like an LLM-based memory, but more human-based.

But I think that will take a little bit of time and training on our end as well. Thank you.

Cool, and then Max and Chan, since you guys have your hands up, I'm happy to stay for like a couple more minutes and answer those questions.

Max, you wanna go first? I can't hear you, Max, if you're talking, unfortunately. Okay, can you hear me now?

Yes, I can. Perfect, sorry. So a couple questions. I've been in, first off, awesome job.

I mean, this was like a full cloud code type of moment, open cloud type of moment. And on X, people have just been going crazy with this, creating groups. And I think it's really rare to have that opportunity. So just kudos to you and the whole SpaceX AI and cursor teams on pulling this off.

Super cool. And in terms of where I run a retail execution business, and we serve some of the biggest food and beverage CPG brands, like the Poppies, Ollipops, Chomps, Siete's. So one of the things that is, we're having a difficult time with, and I think this might come in a lot of service providers, is tribal knowledge documentation. So we have like a tab for everyone to document their SOPs, and even with like record and replay and Kodaks and different things like that.

And right now I have two different technical hires, a senior director of AIBI and an automation engineer that we just hired. And we're all tech outcrop bot, and we're planning on getting this to scale out to the entire team of about 130 people. Amazing.

So we wanna really figure everything out first. So the first big thing is, I noticed the inklings of this, of passive workflow discovery. Is there anything in the product roadmap for passive workflow discovery, where it can monitor clicks, inputs, ways and suggest automations and skills to build for Grokbot based on computer activity? Yeah, I think, so you might have seen it in a couple of the agents where it recommended building an automation.

Like for example, the on-call agent, it was like, hey, my automation is gonna be, you know, polling data dog or major duty. I think another, like a clean slate. The first thing it recommended was like, hey, my automation is gonna be like filtering out any spam emails. So I think there is in the LLM or in the agent itself, there is some preconceived notion of what you might wanna do.

However, I agree with you with the tribal knowledge. I think the more, I mean, and this is true across every AI product, the more context and the more info you give it, the better it will be at actually making a decision or doing something. And so I do think that hopefully as Grokbot gets better, it will be more aware of, you know, passive dynamic workflows. And so it's able to see like, hey, you know, Amrita is often going to Google flights, you know, in, you know, her chief of staff, maybe we should just automatically keep that as a skill.

But I think for now that teach a skill via that video recording is the best way to do it and just feeding it info. So I don't know where your tribal docs, sorry, your tribal knowledge docs live, but like if you can connect to Confluence, to Google Drive, to wherever, giving it that is really important. That's awesome. And one other quick question is the other thing that our team has been absolutely obsessed with right now is we locked into like an annual OpenAI subscription and I'm hoping to switch it to Grokbot and just go all in with you guys.

But one of the things that everyone is obsessed with right now is the voice mode within Kodak where you can orchestrate and people are now setting up skills that, you know, do a pre-flight intelligence and then interview them in the morning and then put tasks, notions that AI can execute and create fan out different tasks to do those. So I was just imagining like, this is such a cool opportunity. Yes, because I know that you guys have a pretty advanced voice mode yourself. Yes.

On the benchmarks. And if you had not only the ability to have like the main orchestrator that you're talking with have that, but also each different bot have its own personality where you can join a group call and almost feel like you're in a room full of team employees where they're all talking back to you with their own voices and personalities and giving their expert advice if you're trying to solve like a really difficult problem, engineering problem-wise or something like that, that would just be like the coolest AGI moment, I think for me. Yeah, and I will say, I think there will be a lot of advancements on that end because we have Grok voice as a part of SpaceX AI, that's gonna be a big part of our product suite. So yeah, I am hoping for that as well.

I think Grok voice is great. It powers a lot of kind of really important pieces of technology that exist already. So yeah, I think that's definitely not far on the roadmap and I'm excited to see how that manifests in the product as well. Awesome.

Thank you so much.

Of course. I think unfortunately I do have to drop folks. I did see two other hands, Chance and one other. If you would like to email me, Nate or Fussell, our emails, I'll actually just put mine in the chat.

It's amrita.x, sorry, at carissa.com, soon to be amrita.x.ai. We're still in the process. But yeah, thank you everyone for joining. I hope this was useful.

The recording will be out soon. Thanks team.

Thank you, Fussell and Nate.

Really appreciate it.

End of transcript. Source: @0xCodila (video) · workshop announcement