Mastra Templates Hackathon, Claude Code is King, and Guests from WorkOS, Recall.network, Clinicamind

What up? All right, we're live. Welcome everybody to AI Agents Hour. I'm

here with Obby as always every Monday. The Admirals of AI. Back to you again.

I see you have your stack of books. Yep. I only have my single copy sitting next to me. O always next to me. You know, I have the benefit of living in the the

publishing warehouse where all the books are. Who knew we were also creating a publishing company as well? Uh here. So, you can get you can get your own copy of the book if you don't have it already.

Mostra.aibook. How was your weekend, man?

Dude, I was in the dungeon all weekend making moves. Didn't even come up for air, huh? Dude, honestly, didn't come up for air. It's one of those like coding holes that I was in. Um, and I'm still in it

actually. So, I I need to come up for air. But I did grow a mustache in the dungeon, so that was a good like use of time. Hey,

dude. I had a crazy beard, right? Because I haven't been shaving and taking care of myself because I'm in the dungeon.

And then, you know, I just made a mustache. I was like, at least something productive happened today in my personal life. Hey, I mean, the mustache is looking good, you know?

Looking like a looking like a real cowboy, you know. Yeah, they have discovered a new drink, dude. Sugar, do Dr. Pepper, strawberries, and

cream. Like, what is this? What is the world coming to? The cream soda one is very good, too. Zero sugar, cream soda, Dr. Pepper. Uh,

hi. This is live as you see this. Kevin, we're excited to have you participate in the hackathon. We're gonna be talking about that in here just a couple seconds. I am as as I did last week, you

know, just drinking my uh sugar-free Red Bull. Not sponsored, unsponsored. Sugar-free Red Bull Celsius either. This is all right. I would uh I would drink it again. But anyway,

um yeah, my weekend I did not I was not in the dungeon, but I did let Claude Code rip. So, I was like it was like a lot of async work, you know. I like sent it on a task, had to do some stuff around the house, came back. Me and me

and Claude Code were big big- time friends over the weekend. I mean, I had had to put Claude Code in his place multiple times. It was going off the rails, but you know, we got it to a good spot. So, we had an intern over the weekend. Yeah. I mean, it it's so wild how

sometimes it's just amazing and then other times you're like, why did you do that? So, it really is like it's like a really smart intern you cannot trust. Yeah. just can't trust it. But it's

gonna amaze you some of the time and you just let it go. I also found, and I hadn't uh checked this before, I'm just going to look at it right now. You can there's a slash there's a slashcost uh feature in this in the CLI when you're just using the CLI mode. So now I know how much this I've been running it all

in one session. So, my current clawed bill over the weekend was $104. Oh, So, I was I was letting it uh letting it really do its thing.

How many time how many uh how many like how many messages was that, you think? I I might tell me. Uh Khalil, you didn't really miss anything yet. We're just getting this thing kicked off. We're gonna start talking about the Monster Hackathon. We're just kind of uh you

know, chatting about our weekends first, but this is live. If you, you know, if you have any questions or anything along the way, please post a comment in whether it's YouTube, whether it's on X, we're on LinkedIn as well. So, drop a comment. All right. So, my total duration says 70 hours.

Oh, Total code changes about 5,000 lines added, 1,300 lines removed. I used uh Claude 35 Haiku and Claude Opus apparently. 668,000 input tokens, 1.1,000 input tokens for Opus, 146,000 output tokens for cloud opus. So

I don't know. Yeah, I don't doesn't say how many messages though, unfortunately. The session was 70 hours though or so total duration running on the API was two hours. Oh, so but my session has been 70 hours

since I I'm assuming that's since I started it. I don't know. I don't know if it's active. Like even if I'm not doing it must be act I've just had it

running. Yeah. So because I've just had the same session running the whole time. Obviously I haven't been actively

doing something with it the whole time. But I've been kind of the profit on your time right there is super awesome. Right. If you spent two

hours versus this thing and it got way more done in two hour. Like I could have done everything that I had it do but it it did it quite a bit better. Yeah. You know it's like Claude's slightly

better designer than I am. you know, hate to admit it, but uh your time is worth more than $100 an hour, you know, let's just say. Yeah. So, that's pretty dope. Yeah. I mean, I think I'm I got to try

to the Claude Max plan because apparently you can cap at 200 and you can go beyond the usage. So, I think I'm going to spend more than that this month easily. I spent 100 over the weekend. So, anyways, we should probably talk about the hackathon, huh?

Yeah, let's talk about it. All right. Well, for those of you that are tuning in, you know, there's quite a few of you, 80 plus of you watching right now. We are announcing our second

or kicking off our second master hackathon. You can always I think we have this somewhere queued up, probably not. Go to master.build if you want to learn a little bit about

it, get registered, and it's all around templates. So, if you were following us last week, we announced master templates, but let's talk about what that is. Let's talk about the hackathon. We'll talk about some of the prizes, who the judges are,

who we have already for sponsors. And if you have questions along the way, please uh please drop them in. So, Beckesh has a question. I mean, maybe you can take

this while I pull up the slides, but what makes your MRA AI unique compared to existing AI frameworks? Are there exclusive features that set MRA AI framework apart? Good question.

Yeah. Um I think that's it's a it's an intricate question because like we're trying to do every agentic primitive that exists um in the community. Um we're doing it purely in Typescript. Um we have our own workflows, memory, etc.

Um it's everything you need to build an AI application. Uh that's what sets sets us apart is actually the community hackathons like this, the people in our discord. That's really what sets us apart. I would actually say it's like the the movement that we're kind of

starting here in Typescript AI. Yeah, I would also say, you know, additional there's of course a lot of features. I think our local dev playground gets a lot of credit for people being able to test and iterate really quickly. So, I think there's a lot a lot around that too. We hear that a lot from from some of our uh

customers. All right, let's uh let's do this. I don't know. All right. So, we're gonna we're gonna

go through this. This is going to tell us all we need to know around uh what the master.build hackathon is this time around. This is our second one. Last time, I think we had over a hundred people submit projects, although there

were some people that submitted multiple. So, you know, it was probably like 70 or 80 that were like unique different teams or individuals submitting. And there were too. What? There were a bunch of bot submissions,

too. Yeah. Or or they just like repeated the submission. Yeah, there's a good handful of like 20 to 30 of those. But yeah, we

got a Well, we have quite a few. And so, let's talk a little bit about the prize pool first of all, like what's in it for everyone who participates. So, we have a whole bunch of different categories. And we'll talk about what those categories are here in just a minute.

But for every category, we'll have a winner and they'll get, you know, $512 Amazon gift card. And we'll also, you know, be launching at some point in the next, you know, couple weeks a community section of our templates library. And we'll show what the MRO AI templates pages today, but we are planning having a community contributed section. So, if you do

contribute a template, even if you're not a winner, you know, we we're just going to pick the ones that we think are good uses of templates that others will benefit from, and we'll highlight quite a few of those in our templates uh page on our website. Yeah, we have a bunch of additional actual like prizes that you can earn or win a

Raspberry Pi, a Nintendo Switch. Everyone who can who submits will get a copy of the book send, you know, ideally sent to you if we can. There are some countries we can't send to, so we'll we'll send you a digital copy. Uh they'll we'll be giving out a copy of the AI engineering book, a mechanical

keyboard, and meeting mute buttons as well. A couple of those. So whole bunch of actual physical prizes.

Uh but we'll talk a little bit. So for those of you who are new some of this some of you this is the first time you know learning about master or thinking about you know joining our hackathon and you're not really sure what master is. So we'll talk about what master is. We already kind of had a good uh pre

Thanksby we'll talk about it what on schedule judging categories and then how to actually submit. So that's the next uh 15 minutes. All right. So we have some questions already.

All right. Jelp says, "Hey, what's up? What's up?" And actual question, can I submit multiple templates for the hackathon? So

we're going to get to that. we you should submit your one best, but if you do have other good templates after the hackathon, you can still contribute them and we'll, you know, we'll likely still promote them and do uh we want people to be sharing these. So whether you can share it, of course, and people can just use it or, you know, we do want to

feature quite a few of them on on the website as well. So yeah, and then Khalil says, "Let's go. Let's go." All right. So, for those of you that don't know, me

and Obby, we're two of the three co-founders of Mastra. So, you know, follow us on X and tune in every week. We're here every Monday. So, we are

if you uh if you've been around, you know, we're here every Monday. But what is MRA? It's an open source AI agent framework for TypeScript. Obby kind of highlighted some of the big features, but we have agents with tools, memories,

tracing, voice capabilities, MCP support. We have a agentic workflows so you can do human in the loop all around a simple API. We have eval. We're just actually improving our evals uh system.

So even more to come on that. We have storage for rag. We have a local development playground. We have agent

networks. We have a whole bunch of different like primitives for building AI agents or AI applications. Any other anything I missed, Obby? We got a lot of and it's all good.

And it's always it's always changing too. always improving because this uh the space is changing. So we, you know, we like to stay on top of it and change along with it. But let's look at the actual master

templates page here. So this is what we just announced last week. And you can see we have a bunch of master templates. We'll go through a few of these real quick and kind of talk about them. But

each of these has, you know, a video. You can users are beginning to expect and you can deploy it instantly to mastercloud if you want to just test it out. But you can also just run this command and you can install this template locally. So what we would like to have happen is, as I mentioned, we

want to have a very rich community section of contributed templates so people can find a good way to get started. Right? A template is one like a starter. It's also like an educational resource for thinking about how you could actually build something like this

and maybe you just learn from it or maybe you build on top of it, right? So, it's a good way to get started. Um, it's also a good way to showcase how you might integrate with different service providers and services. And so, there's

I think there's a ton of integration type templates that would be really nice to have for people to get started. And so this is a good opportunity for you to build something you think would be useful maybe in future projects that you have for others that might want to build similar types of things. So really it's all around like how do we how do we make

it really easy to get started and share these. All right. So we got uh another question here. Can I build and submit templates in the influencer marketing niche?

Sure. Yeah, why not? Yeah. and Shreda jokes, you know, why 512?

Well, why not? Why not? And uh Foundry Studio is just thinking here, thinking about what all the things probably thinking about all the types of templates that could be built.

All right, we're getting the great thing about templates is there's it's the the whole educational component is so understated because a lot of people want to know how to do you know, and a template can just show even if you don't use the template, it shows you what to do and that's tight. And here's another great question from Khalil. Should it be use case focused or tech focused? And actually, it could be

one or the other or potentially both, right? I think there will be a ton of use cases and techfocused templates and I think both are valuable because they sometimes serve different purposes. Right? Use cases probably more

educational but you could still maybe build upon it. Tech focused is definitely more probably starting with and building on it. Right? If you know you're going to use this specific piece of tech and there's a template for it, you can now either learn from that and

kind of copy it into your project or just start building right on top of that foundation. All right. So, let's keep going. Let's let's look through some of the templates that we have just for examples. So, we

have a docs template. You'll notice that every uh each template has a repository. So if if you are going to contribute a template, you'll need a public repo, but it's a good use case if for anyone to want to maybe get, you know, especially if we if we feature your template, we'll make sure we show the author and contri,

you know, potentially contri continue to contribute to it. So we had a docs chatbot so you can chat with uh your documentation. We have a browser agent so you can automate browser actions that's using stage hand. So there's

that one's sick too. I like go ahead dude. Sorry I just I really like this template. This one's tight. Yeah. So that's a cool one and it shows

a little bit of like an integration and obviously it's good. It's kind of use case driven as well. We have this chat with PDF and chat with CSV because those are kind of simple examples where you might want to have a PDF or you have a CSV and you want to have an agent that can chat through that. We have text to SQL so you can use

natural language to query and talk to a database. And then we have deep research which is kind of like uh you know what chat GBT is deep research or like manis or whatever. There's like different deep research tools, but we kind of have a simplified one here that provides a good base or a good starting spot if you're going to build your own. So, those are the existing templates. We

have more to come that will be officially supported, but we would also like to see, and that's what the hackathon's about, a ton of community contributed ones, so we can help anyone who wants to build agents or workflows or agent networks or MCP servers and all that. All right. So, let's continue on and

talk about the hackathon. So, it starts today. Some people I've already seen uh on X. Some people are already started. So, they already shared what they're building. But if you have already,

great. If you haven't, there's still plenty of time. You have all the way till next Friday to submit. So, we'll be in Discord. A lot of people from the team will be in Discord. So if you have questions, go to

the Maestro-Build Discord channel. We'll have a link here that we'll post in a bit. But please join us on Discord and chat with us there. As far as the general schedule, so today

we're doing the opening session that's happening right now. The next, you know, two weeks or till next Friday, we'll be building and hacking and coming up with all kinds of fun templates. And then the on the 8th at 8 am Pacific, that's the deadline for submissions if you want to be considered for prizes. So we

mentioned some of the prizes before, that's when you'll need to submit. And then on the 11th, so that next Monday, that gives us, you know, Friday plus the weekend to kind of go through comb through all the submissions. We'll be live streaming just on this like we're doing right now, the award ceremony. We will be having office hours on Discord

at 10:30 a.m. both Wednesdays. And that'll be we'll just be doing an actual like voice or video hangout if you do

want to come with questions. You don't have to wait for that. Please just use Discord as you go. But we do want to

have some you know if you want to come and chat and hang out and talk 10:30 a.m. Pacific on Wednesdays.

All right. So let's talk about all the categories. We have 12 categories and we have some judges that are going to be helping us judge different categories. You can submit to more than one category. So, you can submit a project

and you can select I believe it's up to two categories that you want to be considered for. You don't need to select uh a few of them. Some of them just like best overall. That's not a category you select. You know, we'll just choose the

best overall. And then the last two are also we'll just choose those. You don't have to submit to those. But the

categories that you'll be able to select from when you submit your project is best MCP server. So if you're actually building an MCP server with MRA, it's judged by Smithery and Smithery sponsoring and they're actually throwing in a switch for the best use of Smithery. So if you build an MCP server and you're using Smithery tools in it,

you will be considered as for that bonus award as as possibly earning that switch or winning that switch. And so Smithery is going to judge the winner of that category. The best use of agent network.

This will be judged by someone on the master team. We have the best use of O will be judged by work OS. The best use of web browsing or web browsers jud or browsing agents will be judged by browser base. The best use of tool providers. These are services like

arcade, composio. There's a number of others, but Arcade is going to be judging that category. So, if you have interesting use cases for tools that you integrate in to your agents or your workflows or your agent networks, Arcade will be judging that. The best rag template will be judged by Chroma. So if

there's interesting uses of rag or rag providers, best productivity, this could be personal productivity like a personal assistant agent or something like that or more of like corporate productivity like a project management system or something. Best coding agent will be judged by someone on the team. Best crypto agent, so if you want to build an agent that interacts with blockchains or

does crypto trading, recall is going to be uh judging that. And we do have work OS and recall coming on today just to talk more generally but also talk a little bit about the hackathon. And then the last two uh Obby and I get to get to pick I get to pick my favorite and Obby will comb through and see if there's any you know hackathon should be fun. So if

you have any funny use cases if you want to just submit something you can still win a prize for being the you know oby's favorite or the funniest. Yep. So, we like to have fun around here. But

a lot of time, a a lot of people asked about examples. Were you gonna say something? No, I just said it's too much fun. Yeah. Yeah, we have a good time. So, any

any examples you want to see, Obby? I have a list here. I can obviously read through it, but I'll I think everyone who's watching can read, but anything that you would want to see? I really want to see like some deep research type stuff that um that like uses agent

network or uses even any of the primitives to do do that type of stuff and then a lot of co-pilot type examples would be really cool because there are a lot of co-pilots out there. So yeah, what about you? Yeah, I think there's a ton of integrations that I would like to see just personally. Personally, one that I

might build if I have time or if someone else beats me to it would be, and this is selfish because I would like this, but we we do these live streams every week. I would like some a rag template that takes YouTube transcripts, you know, chunks it, puts in Vector DB, and then I have an agent that I can talk with, you know, our transcripts from from the show. That would be pretty

cool. That would be pretty cool. And then maybe we expose that on the website or something, you know, and then people can chat with, you know, ask questions about, have we talked to this guest or have we talked about this in the past? What episodes was this on? So, yeah, we got a ton of questions. Maybe we should answer some.

Yep. Let's do it. All right. Uh, let's So, someone says, "Tlates are

so cool." Thanks, Franco. We agree. What about a template that builds

templates? That seems like a code agent. That'd be cool. Pretty cool.

Um, are there any specific guidelines or requirements for the templates? How do we submit it? We're going to be talking about that in just a minute. Uh, we So, we have some of those. We have some docs.

And here's a kind of a general question. As a template, would it be better to have a standalone project or have it inside a NexJS project, for example? So for templates to work and be installable, they need to be standalone, not necessarily embedded with the front end. So we'll be judging based on, you know, just kind of a standalone master

project. Of course, you could take your template and build a front end for it and do all that, but that's not necessarily what we're going to be judging in this hackathon. So the idea is, you know, you could still do that if you want, but we're going to be judging, can we with one command install your template, of course, add some

environment variables, but then interact with the agents and the the workflows in the playground and actually have it do something interesting. That's that's how we're going to judge uh this hackathon. So it's all about uh MRO specific templates, but could be in any of these different types of examples.

And is there a starter boilerplate or sample MRA template? So if you go to maestra.aimplates you can see six examples that we have. It's also they're also in the templates directory in the monor repo they kind of

sync between the two places. So you can check it out there for what kind of a simplified template has. And we do have docs. I'll share those here in a bit. And so to answer Austin questions, do you want to include custom UIs? I it's

not required. So we won't be judging based on that. So, it might be not the best use of your time. Although, if you

wanted to take your template and do something with it, you know, of course, you'd want a UI. How complex does a template need to be? It doesn't necessarily have to be that complex. If it solves a good use case

and it's simple, that can actually be a very good template. Sometimes less is more. So, I don't think there's like a set level of complexity. And arcade is here. That's cool. Khalil,

glad you're here. And then yeah, Austin, yes, templates are kind of focused on the ma master backend server with the playground. So it's not necessarily the full app. Our goal is like then it's a little easier

for people to submit, right? You don't have to build the full app, which could be very complex if you want to build a really immersive UI, a really good UX. And then yes, if you were to win first place in two categories, you could win a gigabyte, but you actually could win a gig one and a half gigabytes because you

could win another category such as funniest or um you know, my favorite or best overall. So you could actually win technically up to three. We pro, you know, it'd have to be really good to win that, but it is possible. Uh yes. Are we encouraged to use these

providers, work OS and Arcade? Yes, you can use them. You don't have to. Uh we we'd encourage you to, you know, you do

what you think is going to be best for a template, but of course, you know, we we do think that they're great and they're going to help you out, but you can either roll your own integrations, you can use these providers, other providers. Ultimately, it needs to be a master template, but you don't have to use specific providers, though it may help you uh get further faster.

All right, that was a lot of questions. Yeah, they're really good questions. Templates are, you know, one part education, other part bootstrapping, getting started, you know. So, those are like kind of criteria.

All right, so let's talk about the prizes. As we said, each category winner is a $512 Amazon gift card. So, a half a gigabyte as someone said, you know, for using that. And then we're going to pick quite a few of them. You know, we're

going to have some like we got to kind of figure out exactly what what qualifies, but we're going to list try to list a lot of community member templates on the website because we want to make these dis easily discoverable. So, just know that, you know, we'll be picking a number of those to actually be featured. And even the ones that maybe

don't make it on the website, we'll still be featuring in social media. We'll be bringing people on if they submit a good template. If you want to come on the live stream, like that's an option. We'll be uh we'll be talking about it over the next month for sure.

We we do have a few kind of early sponsors that already came out and wanted to go full in on helping us out. So, thanks to Recall and Smithery, and we have some more that are likely to be announced soon. So, we'll be talking about them quite a bit throughout the week, but appreciate them for helping with some of the prizes and, you know,

making sure that we can make this even bigger and better than the first hackathon, which was already great. So, as far as what judges are looking for, so just overall appeal, how useful or interesting is it as a template? You know, how educational is it or how well uh does it help someone get started? you know, if there's some creativity, that's

always good. But, you know, kind of like general AI engineering principles. How AI native is it? There's just that's more of just like is it is it a cool AI use case? You know, overall project architecture.

Yeah, obviously needs to be a MRO template. Um, completeness. So, is it a template that would be ready? You could share it and someone could actually use it to either learn from or to get started. And

then, yeah, adherence to the master template guidelines. So there's the docs and we can maybe pull that up real quick and I'll just share where this is at so you can see it. So we have in our docs community section there's a contributing templates and it talks about the requirements uh talks about like template criteria

some quality standards how to you know submit it what required information do you have what the review process is a whole bunch of different details there's probably some other um yeah validation template gallery yeah a lot of documentation you can read there and some other links to I think like the reference and the project structure. So

you can see roughly like how we expect a template to be structured. And the goal is, you know, the reason we want to have some structure around this is we want consistency. That way if someone uses it one template and they're learning from it and they go and try a different template, there's at least like a rough level of consistency between them so

they can understand it quickly. And I think it's just a really good for getting started or educational resources for there to be that level of consistency. You know, I just thought like the more templates that exist, the better LLMs can write. Just say, yeah. Yeah. The more templates we get,

the better we could have a code agent write templates. Yep. That template building template thing. There you go, dude.

Yeah. Um, so question, I'm from India. How's the cash prize distributed? So, it's

going to come across as a gift card. So, it would be you as long as you can access Amazon, you'll we'll be able to get you a gift card. Derek, he's coming on later. I think he wants to switch.

All right, you can win it. You don't you don't get to you only get to judge your category, so you can win a different category. Uh, Venicious, what's up? Thanks for stopping by.

And yeah, let's finish up. Wait, we're not going to ignore this one, dude. We want a Shane bobblehead.

That is a Yeah, I I I didn't want it I didn't want to manifest that into reality. So, we should Who wants to manifest it, dude? I do.

If we do it, we got to do AI agents hour bobbleheads. And then we just put it in the background. Yeah. No one wants my bobblehead. They

want your bobblehead. You know, milk. If they want your bobblehead, it's because you have that sick mustache.

Ah, nice. All right. Okay. So, let's make sure we submit our projects. There's a link.

It's going to be posted in Discord. Um, I'm not going to post it because it's a Google Drive link. You're not going to type that out, but we'll be tweeting about it. It's in our Discord. We'll be mentioning it. So, there is a form to submit

by 8 a.m. to be considered. You you'll need a link to the public repo, kind of

like the demo page or read me, right? Is just the read me. A one minute video. And the video is required because we we

realized last time if you didn't require a video, we got some submissions that were it was hard to understand without someone explaining it. Yeah. Um yeah, so the link is available. We'll

we'll share it in our Discord. So please go to the Monster Build channel for those that are looking for that submission link and we'll be sharing it there and we'll also be sharing it on our X account. And so if you follow Monster on X, we'll post it there. We'll post it in the uh description of this video as well as long as along with the

links to the slides. Um so yeah, one submission per team as well because we figured last time we had a bunch of people that decided to do a you know number of submissions when they would have probably been better chances of winning prizes by doing one better submission. If you do want to build multiple templates, please do. You know,

you can always submit them after the event, but we'll only consider one. So pick your best one to submit for prizes. And we do have raffle prizes. So again, everyone who contributes a template gets

a copy of the book. You know, you'll get a copy of the book, assuming we can send it to you. If you're in a country we can send to, but we'll also be sending out Raspberry Pies, AI engineering book, mechanical keyboards, and those in mute me buttons. Those are just prizes. You

It's like a raffle. So, if you enter, submit a template, you're going to be in the prize pool for a chance to win one of those. All right, any other questions? I know

we're kind of going a little long. We have guests waiting in the wings, so if you do have questions, please drop them in the chat. We've already kind of answered a lot of them as we've gone, but please let us know. If we don't get to them in the live stream, please come on the master build channel in Discord

and we will answer them there. Here are the links. So, our GitHub, our website, we have other events, you can follow me and Obby, but the important one is maybe Discord. That that if you go to that, you'll get to the MRA build channel, I believe, in Discord. So, please go there

and let's build some templates. Let's do it. All right.

What do you think, dude? You ready for see some sweet templates the next couple weeks? I think we're going to get a lot.

I'm stoked. I might reduce the amount of that people ask about certain like patterns and Yeah. And I I'm gonna I'm gonna tease this, but you know, we're going to be doing more than more with templates soon. So, not not going to share exactly what, but these templates are are going to get even more life outside of just

living on the website. And, you know, someone already guessed it in the chat what we're doing, but I'm not going to tell what it is or confirm or deny it. Uh, I didn't see it. I must have missed it. So,

let's see. Let's answer a couple more and then bring the guest on. wondering if crypto category winners would want USDC on chain. Hey, maybe we could maybe do that. We could if make that an option.

That's a good idea. That's a great idea. Um, let's see what else do we got.

MCP client and server templates would be sick. Yes, cool. got this one. Um, it depends. So that's a language level. So it

depends if execution is as fast or not. So yeah, in a lot of cases, if you're just making API calls, it's going to be yeah, very minimally different. All right, Jack's here. What's up, Jack?

We're good. Yo, are we hiring? Are we hiring? H Are we? We're pretty good. We're not I wouldn't say we're actively hiring, but you know,

you never know. If you're cracker, I'm hiring, but if not, I'm not hiring. Yeah. Yeah. You never know. All right,

we are going to move on to the next segment because we are way behind. Yep. But that's okay. Uh, sorry for all the

guests that are coming on later, but we do want to bring on one of the judges of the hackathon and we're going to talk a little bit about the hackathon, but we'll also talk a little bit about just what's going on at work OS. So, we are going to bring on Zack. So, let's bring Zach on. Zack, hey. Thrilled to be here. Thanks so much

for having me. How are you? Great to see you guys. Yeah, great to see you. I don't we

haven't actually uh met virtually besides on I think Slack. So nice to be on a call. Yeah, likewise. Definitely. Yeah, thrilled to be here with you all. And if um audience members didn't draw

the parallel um at the AI engineer world fair recently, my colleague Nick and I gave a a two-hour workshop on how to build um agents in Typescript using Mastra. It was awesome. So much fun. Um,

and we're doing a ton with Mastra internally for internal tooling and just because everybody's real deep in the Genai game these days. So, yeah, thrilled to be here. Um, really enjoyed seeing Master take shape and was like, uh, coming from Pine Cone and having done a ton of like Jupyter notebook maintenance. It's really, really, really nice to be able to see a single rag

pipeline defined in one language. Um, it's quite clean. So, yeah.

um in terms of AI and work OS or Yeah, go ahead. Questions about that? No, just like I'm super flattered right now, you know. Thank you. Oh, cool. Great. So,

there's proof. I I can I can wing you this link real quick at uh Slack Shane and I I had so much fun doing it. We actually wrote of a blog post about it just to share everyone. So there's links there to the YouTube video that shows

the demo of what we did during the workshop and there's also the link to the repo which is open source if people want to use that or yes maybe it's helpful for some template inspiration or something. Yeah, I was already I was already grabbing the YouTube link. But if you have a general blog post that has the YouTube link in it. Yeah, I do. Send that send that here in our chat and

I'll I'll drop it in and we can we can get it on Yeah. to everyone who's who's watching can can grab it. Cool. Let's see. I think it I think I just was able to send it to you and probably chat.

Yeah. So, I'll I'll drop that in. So, it it's going in the YouTube for sure. I think that I don't know if it actually

posts everywhere else, but Cool. Um, awesome. Yeah. And then in terms of just uh yeah, what are we doing

at work in terms of AI? So, you know, everyone is is trying to figure this out. Um, we've been looking at, you know, what can we do to help you deploy uh MCP servers securely super quickly globally around the world. Um, so that's

kind of some fun stuff that I get to hack on during the day. I'll I'll drop you another link here. This is another open source template. It's a Verscell template, but it just shows kind of here's how you can use work OS offkit to

provide the off layer for an MCP server that you deploy on Verscell. Um, and so yeah, my colleagues and I in in dev education kind of have a pretty awesome job. We get to just build stuff with awesome open source projects like master all day and then write about it. So that's dope. That's what I do. Yeah.

What are you seeing out there in the O game like when it comes to like AI agents? Because for like the longest time, no one bothered us about it. Now they are, right? Because I guess people are evolving. What are you seeing

out there? I think that's Yeah, I think it's a good problem that you guys have. It's also because um like master and the playground experience are so awesome that once you kind of get it up and running, you just want to actually publish it. Um that's a great question.

I you know, we're all seeing the MCP aggregators. are seeing like just an an absolute flood of of interest in all these protocols, not just MCP, but A2A, ACP, etc. Um, and I think the answer is that everyone's really trying to figure it out now and at an actual UX and DX level, right? Um, like what are the

right ways for agents to actually uh authenticate? Like how should they represent you later even if they've been running for three days, right? There's like a lot of actual thorny issues to get into when you're actually dealing with a production implementation. Um,

and so we're kind of just trying to cautiously curate and and examine and see what's what's working for folks and what's not and then um provide, you know, just as you guys are with templates as well, excellent kind of like proven out paths. Um, but yeah, certainly an explosion of complexity, certainly an explosion of what's possible, and certainly an explosion of very very exciting uh

opportunities in the future, but certainly tons of complexity as well. Yeah. Yeah.

I saw Michael or you'll call him MG, right? Is that what you all say? Internet internally. We we tend to Yeah.

Yeah. Michael G. Um I saw him give a demo about OOTH 2.1, right? And it really got

me excited about the future of the world. Um and in that like is that like a homegrown work OS thing or is that is he just the one championing it? Like is that made from you guys or is it like a community thing? It's a community. It's community and

spec and we are very plugged into it and close to it and we're constantly watching it like you know to the point we might even like write blog posts in the future about like weekly updates to the spec because they're you know very relevant to us. Um but it's definitely an ecosystem and a community uh effort

to the audience like this 2.1 spec and or something that oh man put me on the spot. I gota let's go look it up together. What are the differences? So I think maybe maybe while you're while you're looking it up because I was at that demo

too because we that was uh AI tinkerers, right? Yeah. Yeah, the tinkerers. Yeah. Um so yeah, it was a it was a cool

demo because it was like how do you authenticate through it through like an MCP server, right? It was like how do you authenticate and then like buy a t-shirt through MCP? Oh yeah. And it was kind of like a okay it unlocks in my mind different

types of experiences and so yeah. Sorry. So here, sorry, that's what you guys are referring to. That's that's MCP

shop. That's also um for for what it's worth, that's live and you can still use it. You can use any MCP client, whether it's cursor or others, and you can buy a shirt and we'll we'll send you that send you that shirt and honor it. It's also

an open source repo. Um and that is actually uh the same place that I I cribbed inspiration for how to do just for cell mcp directly with offkit. Nice. Um yeah, absolutely. But you know the

other thing is like there's just so many there's so many thorny issues too like when you're if you're actually imagining like real world scenarios where you might have multiple agents. Um I was even thinking about this in the context of the AE engineer world fair like what happens when you've got 75 people and

all of them have agents that are all working on their behalf all trying to call Ubers all at the same time. You know all these entrances are are already flooded etc. So I think there's going to be in practice when we actually have hundreds of thousands of agents of you know running in every everyone's cloud um interesting conflicts in the real world

for sure. Do you think agents will have like their own identity like structures you know because I think like right now it seems like generating client IDs and secrets on the fly seems cool and but like maybe you should have just assigned stuff to a particular agent. I don't know. What are your thoughts? This is something that we actively debate every day. I mean, this like from a UX dev's perspective, it's incredibly

desirable to be able to just spin up a project and have an agent defined in it, right? And and some IDP that you you already trust and have that agent be able to kind of dynamically register and get some credentials that are valid. Um, the folks on the security team probably have a different opinion about how cool that is. Uh, especially when it's running long term, especially when we're

trying to answer complex questions such as, uh, but on whose behalf did this agent actually do that? incredibly destructive action. Um, and then how do we actually trace it back in a in a meaningful way? Um, all of these I think

are are fascinating open problems um that are going to yeah that are going to require some kind of community creative building for sure. Um, what do you guys what are you guys thoughts? I mean as framework builders too like how how do you look at it and is it the sense like I haven't shipped a framework before I've used many a framework in my time. How do you guys kind of try to think

about the boundaries of what you want to provide and the types of things that you want to kind of maybe shell out to or provide escape hatches for thirdparty providers? It's a really good question. Shane, you want to go first or Yeah, I can I can take it. I think one of the the most important things is we

know as framework providers that people are going to come with their tools and their stacks for many different reasons, right? Maybe because it's the best at this thing. maybe because it's one person really likes it and so they they have to have this provider uh or they use another they use this provider for

another service and so they need it to work with this provider. So I think you know when you think about having to build a framework you know you you need to make it flexible enough so that third party providers can be inserted at almost every level right whether it's even something that feels like the framework should be able to own it you still have to have escape hashes for

thirdparty providers in and but there's also a lot of things that we don't necessarily want to own. So originally when we started we wanted to own integrations. We thought maybe maybe we could own integrations and then we realized okay well with MCP it's less like we should just rely on that and with all these other providers

that are already going really deep in integrations why why should we we don't want necessarily want to care about that and so I think you know we think of kind of off I think in the same way is we want to provide the tools so you can kind of bring your own o providers but we also know there's a lot of people that care a lot more deeply about O than

we do not that we care, but we're not going as deep. And so I think it's uh that's kind of our general thoughts, I guess. What do you think, Abby? Yeah, like for O in particular, which is a really good example. Like our users

wanted first because Masher creates a server, people want to do HTTP JWT O. And we were like, okay, cool. But then other persons are like, hey, I use work OS, I use Superbase, I use X. So then we

were like okay we should just have an off adapter system which very similar to how Masha works in general everything is a adapter so you can always escape by implementing your own off interface and so that's what we do we are still experimental with this but like we have a work OS one we have a like as whatever superbase etc and if you want to bring your own o solution that's cool and the great thing that for

us as frameworks now that we have an o primitive as O changes in the world, we still have that entry point that we can add features to and then the contract with like for example us and work OS is like super tight. So we'll just always be developing that module together or whatever. Um we didn't do this in the

past. We like learning this now because at Gatsby we did not do like this. Oh, and you got you got your hand burned the first time around and now Oh, fascinating. From experience now. Is that does that

design philosophy extend in some degree to like the way the build configurators work now too in in master? Like I noticed there's like there's tends to be like a a base abstract class and then like hey here's versell. I'm going to gonna gen you the correct JSON just for versell, right? But I know internally like what the actual like markup is for

this. Yeah. Um super cool pattern for sure. Yeah. Yeah. Yeah. We we do that with deployers like

you know Verscell or Netlefi or whatever Cloudflare. We do that we do that with like uh workflow runtimes right you know different whether it's like ingest or temporal or whatever we so the goal is like this we follow this pattern it's kind of our playbook right like we follow this pattern so you know that at almost every level when we can make it

happen you what we'll hopefully have have it for you you can just like easily use or if nothing else you can always roll it yourself and kind of like build your own if you just supply the contract. Yeah. Awesome. I love it. Yeah. It's super fascinating to see like not not having worked on a on a framework product before, it's cool to see like over time

those those learnings, right? It's like I don't want to go down that route again. Also, like what gets cool in a framework is when you have a cloud product that uses it. Like before you'd have to like hijack the user's code to then do whatever

specific things you want to do. But if everything is interfaces, then you can just inject whatever you want. And that's I love that. That's like um I heard last week I think Theo in one of our ads said

I just want to he described requests to people that didn't understand it as like I just want to see my own app's code again. I just want to get back to my own application logic and then like the idea is like there's this pane of glass in front and that handles the login and once you're here you're off right. Yeah. Um love that. Super cool. I I want to there's a there's a question here that I

saw that I just wanted to also make a comment. It reminds me something I wanted to point out. Um the experience that I had personally I think to some degree I'm maybe like your patient zero at work OS because I remember installing master and getting that prompt in the CLI saying uh would you like us to auto inject our MCP server and I said yes please u and that worked the first time

and that that coding session immediately after was one of the first times I experienced like like 98 or 99% uh like oneshot um performance on features being developed because it was consulting the actual syntax of your docs and I I thought that was so fantastic. So this question about like is there a best practice when having ncp server I just wanted to call out is like from the from

the actual like npx cli create app um UX we we all thought that was so terrific and um kind of want to emulate that where possible. Yeah thanks please do I think every should have a doc server now. Yeah that's that that's what we kind of fast followed up with. I built the doc server also from stealing from you guys.

I read your implementation I was like okay let's do the same thing. We can roughly make that same pattern work for us. Um although in our case the the docs have to be kind of siphoned out of a private monor repo first but yeah um that's super cool. Yeah what go ahead and I was say that's why we want to have a template for that so next time it's you

don't have to you don't have to read our code and try to decipher it but you can maybe have a slightly easier starting spot. Yeah. Awesome. Oh that reminds me too. The other question I just I think the answer is yes. But to what degree do you

guys see TypeScript as being essential especially in this Genai world, right? Is that like another benefit of of having the strongly typed language like be able to kind of figure out on its own without you know a couple more human loops some syntax issues? Yeah, like for us well one we're Typescript people so like it's it makes

sense. Yep. We're like it's we're seeing a really a really cool pattern where coding agents are running TSC themselves fixing their own type errors. Sometimes they're hallucinating of course.

Yeah. But because of that type check, you get built in and you get a built-in agent loop of for example in cursor like anytime it can just run tsc or it knows that there are types typescript errors in your file. It'll address them automatically. Um which is there's a benefit. You can add this stuff kind of stuff in Python too, but we're seeing it

just comes out of the box with the language which is nice. Yep, absolutely. Yeah, I'm noticing like cursor kind of pick up those. Oh, I just got four llinter errors with my last edit. Let me go resolve those quickly

before I even talk to you again. Um, definitely a nice added benefit. Yeah. What do you what do you guys think the the endgame is for all the the MCP stuff too? Um, there's the same same

question. I kind of want to talk about how we we also have seen and even support and have resources and examples and docs now for like multiple MCP paths. You know, you can deploy on Cloudflare, you can use their adapter, you can use work OS there in that slot.

Where do you guys all see this this going in? next couple years. That's a tough question. Yeah, you Yeah, pred predicting the future is very hard, especially when we

can't predict what's going to happen next week. Next week. Uh but I I think that I think there's one thing it's pretty clear that MCP is going to be one of the winners for sure like as far as like standards is people are once you get enough of a critical mass I think that MCP in general is you know far and away the most popular standard that's emerged

kind of in this space right there's others that are there and maybe they're complimentary maybe they're competitive but ultimately like MCP is here to stay. I think it's as a spec it's getting better, right? They're continuing to imp improve it and try to make it better.

Um, but I do think that uh as far as like where where this all goes, I don't really think we even know what is going to be accomplished yet by MCP. Like there's people talking like are are agents going to pay through MCP? Like I don't think so. Like I don't really believe that this is going to

happen, but maybe you never know. are uh you know like we at Master of Release is like an MCP course so your agent teaches you the code like that's not a pattern that was ever thought of or supported right it kind of works it kind of doesn't work because it wasn't nec it it doesn't always work perfectly depending on what editors what models you're using

so I think there's a lot of patterns that we don't even haven't even discovered yet so I think it's really hard to predict what's going to what MCP is going to look like in two years it makes it really difficult for us as library authors too because we have to implement the whole spec or at least die trying. But then people don't even use the whole spec, right? Who uses all this

other I don't think I've never seen someone use it yet. Personally, elicitation just came out, right? And I think we're just like we're behind on content even for it, like let alone even using it correctly, let alone even having a best practice opinion for it. Yeah. And I I sit here like you guys and play with the stuff all day long. So

yeah, I do think though that because they're because there's just so many new features and people are getting to the tool call part of MCP and a lot of people just stay there because that's all they needed is just the tool calls, you know. But then I think as more big players, this kind of hap happened in the JS ecosystem too. Once big players start sharing their work on like hey this is

how I use resources to do this useful problem then it'll come into a hype cycle like always it's like oh you're not using MCP resources are you a loser like that and then maybe it'll proliferate more. That's kind of my bet. Yep. That that is interesting. So then also I

guess why why did MCP my feeling and perception is that MCP's caught fire sooner um than A2A which I guess it's because it's faster to show value like as a single developer I can sit on my machine and get something working with MCP and now I can see that like my favorite LLM friend has capability that it didn't have before. Right. Exactly.

Do you think that's that's the reason or when we were talking to people about A to A they're like dude like one dude was like I barely am building my first agent. Like why do I need another one? you know like I'm not in a multi- aent network that I'm gonna protocol but maybe as but then what we realize all the time and Shane can attest our users are evolving every day as they are

learning more stuff so maybe not today they don't need a but maybe eventually they do and that's what makes it challenge challenging for us to both make sure we support all these things but you have to predict the future of like where are they evolving to next you know yeah I I do think that I do think that M like the problem with A to A is it it is

like kind of agent to agent but I do think a lot of people are just saying why why isn't that just like a tool that I can give an agent that talks to another agent through MCP so I think there are there's a lot of overlap there and I think because of I think MCP came on at the right time like people were finally starting to ready around like

how do I give tools to agents I think that was a common problem people were having and so I think it hit the nerve at the right time by we knew someone actually in YC that was kind of coming up with something similar that obviously didn't catch fire because they didn't have the, you know, social clout that

Enthropic had for releasing it. So it's like there's certain like that was my question too. It's got to be a perfect storm of like you got to have the distribution and the reach also have the timing right and obviously like Google with ADA had the had the distribution but I think the timing wasn't quite right and also now it's like is it owned by like Mozilla or

something like transfer the ownership like I don't know what's that foundation Linux foundation. Yeah. Yeah that's right. So they moved to the Linux Foundation. So I'm just like I don't know how well that's going to be. Maybe

people will evolve and get there, but maybe they're just gonna make MCP work for that use case. That's my that's kind of my my inclination. There's one other there's um coming from from Pinecom before, there's one other place where I think you guys are probably a couple um definitely ahead of the game too is an eval. When I saw that

eval was already baked in as a first class, you know, so not just help me master can not just help me set up a pipeline and run it but also understand from pipeline run to pipeline run is there degeneration is the you know the pipeline healthy essentially in my is there some kind of like skew towards

like a a case I don't want to go to and that's fairly advanced like even though anthropic and open AI talk about that being a best practice of actually deploying agents in the enterprise um for for customers like most people that I work with are not already thinking about eval and they definitely don't already have it baked into the product,

but it is the kind of thing that once you do get up in production, it's like now you immediately need to go and you're going to want to reach for that. Um, so yeah, thought that was really interesting to see. Um, that was another sign of maturity for me because it felt to me like, oh, this is coming from folks who have have suffered over pipelines before um or have, you know,

at least burned their hands. Yeah, for sure. Uh we're kind of we're actually re we redid our whole eval framework internally or like externally too and we're like doubling down on evals as like to share more primitives because it's funny like we kind of like the framework reflects like our collective knowledge and we're like learning evals right now and now better you know so that's

yeah that's so cool that's it's so fascinating to think of it that It's like of course obviously it is like it's the product of like human engineers working together but you don't explicitly think of it that way all the time. Yeah. Um I'm curious the degree to which both of you felt like oh wow finally this pattern of agents is available and feasible in today in this era and now I

can reach for it. Um maybe the better question is like to what degree did you feel the lack and the missing of agents before? You're like, if only LLMs were at the point where they could do this, or do you feel like everyone is like, hey, this is agents year whether you like it or not. So now it's time for agents. We tried building our own before MRA. We

tried building AI futures and like I don't know, hindsight 2020, I wish we had MRA back then because if we did then we wouldn't have built MRA or whatever if you know what I mean like wouldn't the need to even go down this path if like this existed back then. I I I do think though there's a combination though like yes the tooling sucked but

also the models were not as good. So it's like a it's like there has to be is like the tools get better, the models get better and eventually you get to a certain point where new use cases of agents are unlocked, right? And I still think I think everyone would probably agree there's a lot of things that agents are not reliable enough for. But every time

there's new model updates, every time the tooling gets a little bit better to help you like connect these different systems together, you can start to think about, okay, there's more use cases now that would maybe still be pushing the edge, but might be feasible. And then the models get a little better and they become more feasible. It's like now we can maybe try to do something a little bit, you know, outside of that. So I I

don't think there was like a tipping point for me. I think it was like when we started it, we're like, okay, there just needs to be better tooling around these LLMs. And then it's like the LM kept getting better. our tools kept getting better and I think we could

start to see a little further into like what could be possible if these things if it keeps going with this trajectory like more and more things are going to start to open up and I think a lot of people especially those that are not kind of in the space like thinking about it are going to be surprised at how many like potential uh unlocks are going to happen as as

these things continue to to start to like even iteratively improve right even if we didn't get a major model like release I think we can still do a lot of great That's so that reminds me too vers um AI SDK which I I loved and found that you made a design decision to build on top of that and which also thought was a

brilliant idea. I remember it was like a year ago or so or a year and a half ago like I had built a rag pipeline in into my site just based on my own writing using their stuff and I remember the day that they announced hey the 40 model has dropped and so that was a in my case a one character change to go from four to 40 push that PR through in two minutes

and then the entire pipeline was like stepwise upgraded with better intelligence and I made zero application changes so that was yeah that that's um that's another reason I want to be on the Verscelli I SDK Yeah, for sure. Easy for model routing. Yeah, that's great. Definitely. Um, do you guys I have one more question. Do you guys have a sense that in the near future um maybe it's

whatever it's transformer 2.0 or some other architecture there's some ability of models to uh continue generating quite quickly but also generate with like let's say higher accuracy. So we get to the point where I ask for you know um whatever next.js JS application of a varying complexity and I get like

99% of what I was expecting and I get it in like under a minute. Yeah. Um how do you do you guys think that that world's coming and what do you think that that changes from a Devox perspective? We do this session on the show where builders do ML where we have

our ML friends come on and teach us And what we learned last Professor Andy by the way if you're watching thank you m so much professor Andy but what professor Andy told us was like when you're trying to get accuracy you are you need a bunch of data points and you're training you're doing reinforcement on a model. Um, and then you're essentially what what accuracy means is the the model is calling tool

calls without having to reason about them. It's just and then like you're just training it to then execute uh tool calls given certain eval cases and stuff. So like if he if if we were talking to him right now, he'd be super bullish on the future of like accurate stuff, but he doesn't think it's going

to be from a general model. You're going to have to do it yourself. And I guess that's kind of what the business is kind of built on. So find in the sense of

find like taking a base a foundation model and fine-tuning it to my specific syntax or use case which I also then so then in that world we would expect frameworks maybe uh or existing frameworks to have incredible tools for like you know distilling a model from some other model or quantizing it getting it onto my Android device super quickly so I can go walk in the woods and talk with it offline or

whatever. Exactly. Um I think like model operations, let's call it model MLOps ML MLOps is like it's like secret like a secret fraternity, right? That we are not right

now. It's beyond secret. It's like you you watch that pietorch documentary. It's pretty it's pretty mind-blowing. Yeah,

they're like a mafia that we need to like infiltrate an Illuminati of bringing the knowledge and share it with everybody. Yeah, I have a couple of those textbooks behind me. They're super fascinating. Um,

yeah. I I mean, so I I I will uh take the flip side of this just a little bit. I'll pour some cold water on on expectations. I don't have any inside knowledge. Uh G John says he thinks it's

coming next month. You might be right. I will take the other side of that bet though. And I think that it seems to me

that the hype has now kind of superseded the progress. Like I think we're model implements, right? But we're having to find new benchmarks because eventually like the benchmarks like once you get past knowledge it's like you have to do more than just be able to recite facts or answer questions. You have to be able to take multiple steps and then like LLMs

can't reason, right? It's just next token prediction. And so it's kind of amazing how far that's taken us. I don't

know if that's going to take it as far as people think. Maybe some people think with some of these other techniques it can. I think it, you know, I think it'll continue to improve. I don't know if I'll see this if we'll see this big step change. I don't know if the hype's going

to live up, you know, if GPT5 is going to be this, you know, massive step change or not. I kind of doubt it, but I could I would love to be wrong because I think that would unlock even more capabilities. But, you know, I I believe the hype, but not in the same maybe time scale of what some of the uh really big

advocates of what's coming are thinking. So, we will see. I don't have no inside knowledge, but you know, we we'll see who ends up being right in this. I think, you know, eventually we'll get there. I just don't know how how soon.

Super cool. Well, Zach, it was great having you on. It was a pleasure to be here. Pleasure

speaking with you both. Thank you so much. Yeah, you you'll if you ever have a cool demo or anything you want to show, come back on again. Happy to have you like

showcase cool stuff with Maestro, with work OS. Hopefully, we'll have some really cool uh templates for you to judge. Maybe some that use use work OS, but hopefully a lot that use interesting use cases with Oth. And yeah, excited to have you all helping us judge the the hackathon.

Terrific. Thanks so much for having us. Really really pleased to be here. Thanks for coming. Yeah, see you. See you,

dude. That guy is cool. Yeah. Yeah. I never met him before, but besides us is super cool. Everyone I've

met Yeah, we we've seen him at some meetups. Definitely always Yeah. Honestly, just like their level of excitement for how big of a company they are around around this is just really cool. Yeah. And they're like, you know, from Michael like going to the meetups and talking and and being part of it to

they're throwing like MCP 2.0 night or whatever too as well. Like that's a big big event, dude. Tool. Yeah. MCP night. That's gonna be wild. Yeah. What if that's like the biggest party in

San Francisco? Yeah. I'm not going to be able to be there. I'm really really upset that I'm not gonna be able to be there. But

you're I'm assuming you're going, right? No, I'm going to that Defcon. Oh, yeah. Be there. So Sam Sam will be

there. We will have people from Mastra. Sam and Ashwin will be there. We will have a representative there.

Yes. So if you're going to MCP night, you'll see someone. But we're already behind. So let's keep moving. So we're

taking a little bit of a divergence from talking just about the hackathon. We wanted to bring on someone who's building with MRAA and talk a little bit about how uh agents and AI can be used in medicine and uh see kind of the cool things that maybe David's been working on. So, we're gonna bring David on to the show. Hello.

What's up, David? Hey. Hey. How's it going? Thanks for having me. Sorry for my voice. I'm a little bit

sick. I woke up sick today, but um happy to be here. Well, I'm glad you were able to uh you're still able to come, but hopefully you feel better soon. Thanks so much, man. Yeah, tell us a little bit about uh

yourself, a little bit about what you're working on. And yeah, I know people are always interested in like use case for how are people actually using master or even just agents in general. And so I think just talking through that would be really cool. Awesome. Yeah. So my name is David. I've been coding since I'm about eight years old. Started coding on commer 64. Sold

two companies very early in my life. Um once a unicorn at this point and uh I ended up falling into addiction um in my early 30s. Um ended up homeless the whole nine yards. alo those guys in the streets that you see here in San

Francisco but I was in New York and uh as a result of that I got back on my feet eventually but then I went to rehab and I was inspired to build in recovery my last company that was all focused on datadriven care I've been building ML since way before AI was sexy um with computer vision stuff and so on but um now with you know we exited that about a

year and a half ago now we're building clinical mind which is all focused on healthcare AI we see a massive opportunity to completely change the paradigm of how doctors and clinicians and patients communicate in healthcare, really focused on conversational AI. Um, and really thinking rethinking the whole

way that that that the medical record works uh from from the standpoint of a communication platform rather than a documentation platform. So anyway, we're doing a lot of cool stuff. I met Sam from Mastra just recently. That's how we

started using you guys. He's phenomenal. I love him. Um, and also Ashwin. Ashwin and I have so much in common. So we we

actually were at at the hackathon yesterday. um just jamming with uh with a bunch of really interesting projects that uh that were used in Ma. So yeah, happy to be here. That's my background in a nutshell. But um nice to meet you

Abdi. I haven't met you yet. Yeah, awesome, dude. Yeah, excited excited to hear, you know,

that you you found us and you're you're using MRA to build some of the cool stuff that you're working on. Uh I guess can you tell us Yeah. Can you tell us a little bit about what is Clinico mind and how does it work? You know, how

Yeah. So, Clinical Mind in a nutshell, h it's so hard to explain it. You know, we we're actually going to be uh applying to YC. Uh your same uh partner um Kustaf actually told us to please apply and we

didn't because we I went to New York and we ended up selling three contracts before we have a product. So, now we're building. We're actually in build mode right now, but we're already have about a 100,000 in ARR, which is wild. But um

we we're basically using MaRA currently to orchestrate a lot of workflows in healthcare. I'll show you some examples. I'll actually share my screen. Um let me see if I can Yeah. Yeah. This will be

the best I think. Show show it visually. So yeah, sorry.

Live demos. Live demos. Well, actually not not so much a live demo, but uh I'll just kind of show you guys what we're working on.

I didn't come prepared for a live demo today, but we'll definitely do one next time. Uh give me one second. So, how do I share my screen, Shane? So, there there should be a share button kind of in the bottom middle of of that.

Or if you press the H key on your keyboard, I think it I see it. I think I tried to share, but it didn't allow me. Let's see if I I'll give it another shot. Yeah, you should have permission. Okay, I'll try it again. Uh, you know

what? I'm going to restart the browser. I might lose you for a second, but I'll be right back in. Okay. Yeah. Yeah, no worries. We'll see

you in a bit. So, this is a great opportunity while we get our guest back on. This is a live uh live show. So, please if you have any

comments, questions, you can use the chat, whether you're on uh X, whether you're on YouTube, whether you're on LinkedIn, please uh drop us a message. We try to answer a lot of questions. We will be getting another guest on here in probably, I don't know, five or 10 minutes. But awesome. Uh, David, you're back.

Yeah, I'm back. Okay, awesome. So, I'm gonna share my screen. Let me know when you Oh, screen sharing is not allowed. I

got a message on the uh on this studio reream. That is No. Yeah. I don't know. I don't know why. We've had many guests

share their screen before. So, I'm curious what's going on. Let me I can build you a tech company, but I cannot use conferencing software. Yeah, exactly.

Um, what if I try just to like screen share? Yeah, it says screen sharing not allowed. Unable to share screen. Click help center. Let me see. I don't know if it's my permissions, but

it seems to be on your end. Let me see. Maybe we changed something that not on purpose. So, we're just we're just building a framework over here. We're not, you know, we are uh we're we're hobbyist

live streamers. We don't we kind of we joke we don't really know what we're doing. We just show up every Monday.

That's awesome. Same thing here. I just show up really excited every morning to to just build cool stuff. So um anyway,

one of the use cases we're building uh with MRA is actual call center for healthcare. So we have uh we're actually using Bappy for the conversational side which I know that you guys are starting to partner with. Um, but Bappy's limitations kind of hit really quickly because we we need a very deterministic workflow and that's what I love that you

guys ended up using. I love David Piano. I love that you use XATE under the hood. Um, been using Xate forever. That was

one of the reasons why it became so appealing to us that there's a way to actually make agents a little more deterministic because, you know, by definition they're not. And so uh using the basically mapping out the workflows that this hospital wants has been just phenomenally useful. Um and we're going

to continue doing that. Also for example many use cases in medicine are very algorithmic. So there's like for example something called up-to-date which is what all clinicians and doctors use inside the hospital setting to see okay if somebody comes in with let's say a seizure this is the exact path I'm going

to follow. So we're kind of giving that to our agents in many ways so that they actually follow very specific path. They are essentially helping doctors become AI native. We're not replacing doctors.

We're not replacing people. We're very focused on enabling um these clinicians to actually be able to do more with AI. There's a lot of noise out there. I think there's a lot of um there's a big push for scribes which are really I I

think a lot of the healthcare AI startups are missing the mark. Uh we're really focusing on making healthcare more human again through the use of AI ironically, right? So that's kind of how we're using Masa currently especially on the workflow side and um the eval piece. We I I I I heard the the last

conversation. It's such an important thing. That's actually one of the one of the winners of the hackathon yesterday was using Mastra and specifically like love the eval component of it. Um we're

using weave actually. So we're tying it together and that's another partnership that I'm trying to enable with you guys with wits and biases and and MRA. Hopefully that that also pans out. Um but using weave we're able to really be

uh sorry there's a question there. Does your agent architecture look like inside clinical mind? using single. So we're

using actually multi multiple agent orchestration. Um we we're not really focused on creating like these massive multi- aent settings because then you end up with like the same thing that we end up with microservices, right? Like just bad architecture just leads to bad results. So but we where where it makes sense and we're actually giving each of our agents a persona. So we are like you

guess we're also talking about this earlier. We are actually creating a very specific persona around an agent. Let's say it's a call center agent or it's a clinical assistant agent or it's a u let's say a lab expert agent that all they do is that and the reason we're doing that is because every provider that we work with we're actually

building their own model for that provider. So we're find we're actually creating a a graph neural network for every single provider provider is a hospital a clinic or even a uh an independent doctor. By doing that, we're able to then feed AI into that model and it just like it's superpowered, right? So like it understands where it lives,

what it does, you know, what's the resources I have available. So beyond MCP, we're really creating a way for the agents to be fully aware of of what they're of where they're where they're living, which is really really important for healthcare. And it really creates a lot of safeguards as well. So like when um all right so let's say

someone starts using your product are they like are you ingesting all their medical records and stuff to then be available like what's how is the surface area of this agent exposed to the user? Yeah. So this that's a really good question. So we're actually using uh Apollo GraphQL on the back end. So we're

big GraphQL people. keynote at a couple GraphQL conferences in the past and uh we are using that data layer to limit the agents access to data. So we we're kind of circumventing the whole MCP security issue uh by actually having our data controls on at the data layer, right? Um and then what we're doing for

for the agents to to access that data, there's an there's also an authentication layer. We actually are experimenting now actually with someone that we met uh recently who's really solved some of these very big issues of of Asian authentication. Uh he's a cryptography guy. He's actually going to be at Defcon, too. I might actually be going there, Abdi, so I might run into you. I've been going to I've been going

to Defcon forever. I I uh I I always uh tell my story that I I got kicked out of high school for hacking the school district, but I've been going to Defcon for for a long long time. Um, yeah. So,

I was just around. But anyway, um, I mean, I was playing around and it was solving puzzles more than anything. It wasn't even malicious. Um anyway, so uh this guy that we met that's going to

be presenting this at Defcon has he's just a genius from from Berkeley uh a professor there at Berkeley and he's created a way to take um essentially holographic uh encryption and make it real like a holographic manifold is collapsed into entropy and it's just like this insane amazing way of creating

an extremely secure keyless system actually which is really what we need for for agents and and we can maybe even have him in here uh one of these days. I think he'd be a great person to explain what he's built because it's just phenomenal. It's called Enigma. Um anyway, uh so we're using that to to basically uh

enable our agents to access certain medical records, right? So um it's very important. Privacy is extremely important in healthcare. So we are always making sure that whoever is

accessing anything whether it's human or AI is part of that authentication pool and it's also within the audit trail of everything. So again evals and and just observability in general is is such an important thing and I'm glad that it is that it's actually built into to MRA. Yeah. Thanks. Yeah. Yeah.

What's the challenging part of working in healthcare? Regulation. Um, I think a lot of HIPPA older Yeah, HIPPA. HIPPA is a a joke, though. I mean, security-wise, anyone that every every single person that's

going to Defcon can beat HIPPA in their sleep. But I think, you know, to be honest, it's more so HIPPA was created by by lobbyists essentially. So, they they didn't really they didn't even tap they didn't have the the decency to tap the security community to be like, "Hey, does this look right?" Uh GDPR though in in Europe, that's phenomenal. like

they've actually even focused on on the biggest uh attack vector which is the human you know the social engineering side is actually covered on GDPR where it's not really covered in HIPPA but really I would say the biggest challenges of building in healthcare I mean there's way better ways to make money you have to be really passionate

to be in this space uh because you will be you know just facing crazy long sales cycles incredibly like long long um compliance calls with with hospitals and so on. So I I would say like the biggest challenges are the regulation the the long sales cycles and and also adoption adoption of people that don't want technology historically right they just want to come in do their job they don't

want to be bothered they don't and that's why like for us with with clinic mind we're so focused on making the UI conversational so it's not even there like this the ideal way like my last company in recovery we were launching um we brought design thinking into the into the healthcare space. So we basically during our deployment process we were involving the clinicians all the

frontline workers were part of the development process. So that was their button that was their idea. So they were listened to and and and it actually made up made the product so much better because we were literally involving our customer from day one in the buildout. Um, that's one hack. Anyone that's doing anything in enterprise, that's a great

way to to to get actual adoption post rollout because you have obviously the contracting long sales cycles, then you have the deployment sales uh cycles, and then you have the adoption. And if you don't get adopted, then you're pretty much like done, right? So, there's a lot a lot of challenges. Yeah. But it's fun and we're doing something that's really important.

I'm seeing a bunch of questions in here. Yeah, let's answer a couple questions and then uh we'll we'll let you get on with the the other stuff that you have and we'll get get our next guest on. But a couple questions. Absolutely. Are you using a monor repo? How easy was

it to use MRA? You know, asking just in general about Ma. So, we're not a monor repo right now. We

will be soon. Actually, I think after you you interviewed one of my really good friends, uh Andre uh from Perplexity. He's uh he's actually building our our our monor repo framework now. uh he's an expert in

that. Um so we will be a monor repo soon. Uh for now we're not. So I can't really answer the how easy it was to use

that inside of MRA. Hopefully it'll be I'm sure Andre will figure it out. And we we do have monor repo support. Uh

awesome. Perfect. But there's there's always bundling issues with monor repos and stuff. So if you so if you do run into something, if

you're if you're listening to this and you are using a monor repo and you're running into problems, please let us know in Discord. We're we're continually improving our monor repo support, but it's an ongoing battle. GitHub issue. Yeah, even better. That's that's the best way to do it. We actually are going to use NX. So maybe we should even just create an NX

function for for deploying U MRO. We can just contribute that to the open source uh part of this. Uh which would be awesome for anyone building model repos with NX. Um the other question there that was actually a

really good one. So how are tools defined in your agents APIs microservices tied directly into the HR logic? So I really I so wish we could share my screen because I have something really cool that we built at the YC hackathon. the MCP hackathon. That's actually how we met Gustaf. Um I was

downstairs in the basement having a call with the guys from ApolographQL who were helping us with the MCP server that they had they had just launched two days prior. It was a Saturday I think and he the the guy that coded the actual MCP server came on. He was making uh pe uh peanut butter and jelly sandwiches with for his daughter and he's just like

coding with me and that's just the kind of environment that they that they are. Uh but anyway um I uh we we are using GraphQL to the max. Basically the only thing that we're exposing are actually queries and mutations at the end of the day. We are using resources from the MCP server especially for subscriptions

because it's really useful. Um but we are still kind of in the early days kind of just using operations for now like pretty much everyone else but there's so much more to the MCP protocol that people should 100% take advantage of. Um but for now we are basically just exposing certain queries and mutations that we want to expose and those are the

ones that then um have the actual API connections to the EHR. Um if that answers your question Babex clinical mind tech stack. Oh that's a fun one. So um yeah so all no so all

Node.js obviously the back end. We're we don't believe in Python back ends. I love Python but not for not for production. Um, it's great for it's

great for ML and my god my my my colleagues are going to kill me for saying that. But I just I think you know we've all proven that's why I love MRA and and the the premise of Mostra is that you have you're you're giving access to you know millions of developers to finally start building AI and I think that's just phenomenal. I commend you guys for doing that. So text uh we're nojs we have an apolographql um

layer which is all of our queries and mutations to our back end. We have a bunch of connectors that we're building for each uh EHR. So we have one for Epic, we have one for Nextech, a couple of other ones that are that are smaller.

Um but basically we're going to be building connectors or actually more so adapters for every single um um EHR that's out there as we start deploying and and getting more clients. Um then above that, we're we're actually building all in React Native. So React Native web and React Native obviously for iOS and and and native and uh and

and Android. Um and then uh in terms of the AI stack, Maestra obviously for agent orchestration. We're using we actually were building our own conversational pipelines with Pipcat, but we're now using Vappy just for you know we needed something that works that's out of the box. That's awesome.

Um we were using 11 Labs in the beginning but that we hit very fast the limitations there. It's actually 11 Labs is great for MVPs for sales but um for actual production it's it's a little bit too too restrictive. So for now it's Vappy. We will eventually I think build our own um conversational pipelines

because it's just so important. And then we have MongoDB as as our as our document store and then we're using also Neoforj for all of our graph neural network stuff and PG pi for the graph neural networks. Yeah. Cool. That was a deep dive and we will definitely

I want to show you guys. Yeah. Because it's Yeah, we we'll bring you back on. So, we'll figure out the screen share issue in a couple weeks or, you know, we'll

bring you on and let's have you do an actual demo because I think Yeah, I'd love to. I'd love to. I think people would like to see everyone always likes to see real projects, right? Real things built. It kind of opens their mind of what's possible, especially because we have a

lot of people watching this that different stages. Some of them are using MRA in production and some are just evaluating, you know, should I use MRA? And I think stuff like this helps them learn what's possible. Yeah. Yeah. Yeah. Um so uh we actually

are open sourcing a lot of what we're building. So much of it will be available on our GitHub repo soon. So if anybody wants to maybe what we can do Shane is we can even create like an example project for medicine or for healthcare using MRA for some of the a template. We make it a template. Oh what are those? I didn't I didn't

catch that. So yeah so we are having our MARA hackathon our templates hackathon. So basically a template is kind of like a starter. It's kind of like a either an educational resource or an actual like

getting started. So, it could be a collection of agents or workflows and you could just, you know, contribute a template. You can win some cool prizes, but also just, you know, we'll we'll list a good number of templates from that are from the community on our website as well. So, people who do contribute templates can kind of get

those get those out there. That's super cool. Yeah, I'll love to contribute something to that. Um, and uh

yeah, this has been awesome. Thanks for having me. I'm sorry that I'm a little sick and I apologize for not being able to share my screen stuff with Ashwin. Uh he was excited to share it too, but we'll do it next time. Yeah, no, no worries. We uh we're happy

you were able to come on and talk a little bit about, you know, agents and AI and medicine. And we didn't even get to all the questions. There were so many good questions that that came through. So, we will uh we'll have you back on again and we'll maybe do an even deeper

dive. So, appreciate you coming on. Yeah. Have a good one.

All right. We are way behind everyone, but we're gonna continue trekking on. I I told Abby earlier this week, I was like, you got to buckle up. This one's going to be a long long show because we

just have so many uh cool things, cool guests, uh cool things to talk about. There's We haven't got to the news yet. We got So, buckle up. Uh hopefully you have your Celsius uh and some, you know, Cheers.

And with that, let's continue trekking. I'm excited for this one. So, we're gonna bring on Derek from Recall. He's

gonna tell us what Recall is. Gonna talk about how Recall's one of the hackathon judges in a in the crypto category. But Derek, welcome. Welcome. Hey, thanks for having me, guys. Are you

sponsored by Celsius or what? No, we wish though. Not yet. Just rep it. You rep it hard. Yeah,

exactly. Not yet. Let's do the shot. If the people watching this make it happen, you know, we can't we can only

do our part. We need people helping us. That's right, everyone. Get out there, quote, tweet. Um, thanks for having me,

guys. Yeah, where do you want to start? Yeah, I guess you know, maybe just a little bit about you and then a little bit about recall because I think, you know, our worlds overlap quite a bit, but I'm sure there are people that don't even know what recall is. Yeah. In the So, I'll do me, I'll do recall, I'll do overlap. Uh, hi, my name

is Derek. I used to be a high school physics teacher, only child syndrome. definitely wanted to be like a Bill Nye Carl Sean type which is how after burning out from public education uh I ended up as a Devril so I'm dev lead at recall network um and that's where I get to explain to people how to build AI

agents book mentioned um thank you by the way I had to hunt you down Shane for this but I was nevertheless he persisted um yeah so at recall yeah I guess the super short version is like we're trying to be Google's page rank, but for the internet of agents, because I mean, I keep on top of this stuff. I know you guys do. It's literally master's job to make it easier

for there to be more and more agents every day. And so, that raises an important question, which is like, how do we find the good ones? It turns out there's like a lot of them and they're not all good. Like, I literally am using

every screen recorder software, every email writer. I'm like running my my test suites through different models. um to say nothing about like the productized versions and it's just like hard to keep up. So recall is trying to

do this um agent rank type um head-to-head competition for various capabilities. So like a chess ELO score but for agents. So right now we're focused on trading competitions. So

that's like trading agents, hedge fund agents, that kind of thing. Uh we're cryptonative. So, um, it's a very natural kind of, uh, go to market strategy for us, people that are like, I want to prove that my agent can like make people money. And in a world where you let the agents, um, sort of battle

it out and then figure out who's best and have this dynamic rank over time, um, the the work OS guy, he was talking about how it was just like a one character swap in his codebase to go to the next um, model when 40 came out. It was like, what if we could do even better? Because like open router already

lets you kind of toggle between different models and stuff. What if you could just do like, you know, import recall SDK and then recall whoever the best email agent is this month? I don't want to have to like worry about that cognitive like overhead. So, that's what we're going for. We're just trying to add this discoverability layer,

verifiability layer of like, are you good at what you say you do so that people can find you? And speaking directly to the uh agent developers watching this, wouldn't it be nice if you could spend all your time just making your product good and not have to invest so much time over on the marketing hype side? So yeah, that's

kind of the world we swim in. The overlap here is of course um I shall master all day long. Like we just wrapped our uh online hackathon this past month. Thank you so much Shane for coming. Um a lot of people were like I

don't know where to start. What do you recommend? And I was like easy master. We have um we have a guide that um I I

put a link in the chat if you want to throw that up for folks. It's literally that easy of like just grab Mastra, go through a couple extra steps. We're from fresh CLI. This is not some fork that's going to get, you know, behind main. You

just go npxmra and then add a couple steps to hook up to our API and then boom, you're in. So we're we're not like here to solve the crypto trading problem. We're here to solve the entire discovery discoverability problem. Um,

but like I'm I'm teacher man. Like I like inspiring people. I I love the 50 people watching this right now of being like, I want to make an agent. What do I do? It's like, well, good news. Typescript is eating the world. So,

we're here. Do you guys get like because you're crypto native, do you like get any into trouble with any of that type of stuff or like do you I don't know what the trouble is. We've had early on we had like crypto scammers and all this stuff try to infiltrate us. And does that happen more often? Oh, you mean like um fishing or what?

Well, there's fishing. There's people on Twitter saying that we released a coin and Yeah. Oh, yeah. Yeah. That's just another day in our industry. But I will say um you know one of the reasons that

crypto and AI agents in particular are kind of a match made in heaven is it's like an internet native way of doing important things. Like all you need is a private key to be able to like custody funds, interact with smart contracts, make certain covenants even crosschain. So yeah, we've definitely seen like a

boom um you know with with virtuals and and the you know base network and Salana network a lot of these AI agents um it's pretty cool. I mean, like I I'll whip up the master of local host, the the 411, the UI, and just like throw in an API key and just be like, "All right, what tools do you need to be good?" And it's like, "Uh, go give me these." And

I'm like, "All right." That's my favorite mode of interacting with AI, by the way, is like, you know, that moment when you're coding and you shift from like, "Here's what I need you to do next" to like, "Can you teach me how to think about what I should be telling you to do next?" And it becomes a two-way. So, I'm I feel that way about my agents, too, where if you want, I can screen

share like a demo my colleague threw up um with Mastra last week. Yeah. And to get there, it was like, "What features do you think you should have?" And it was like, "Well, you should give me a Reddit API key and a Twitter API key and like a news API

key so that I can actually scrape for more data sources." And it was like, "Yes, sir. Let me go get you those keys, man." So I I think the more we treat agents as um

you know people spirits to use the to use the term is like you know ask them questions like treat them as like colleagues and it's like they are fully capable of answering those things and surprising you. Yeah. I like the people spirits uh way of like thinking about it because then it changes how you interact with the AI too if you think of it as a people spirit versus just a task

execution list. Yeah, I'll I'll watch some of my uh colleagues little screen share and like as soon as it starts hallucinating or whatever, they start yelling at it. I'm like, "You got to go to therapy, dude." Like, I know for a fact that the system prompt, at least on

Claude's side, but I'm sure for all of them, is like match the user's tone over time. Like, if you've ever been in those sessions where you're like, "LFG, fire emoji." Then it starts firing back like to the moon, rocket ship, rocket ship.

Like, it'll it'll match your vibe. And so, it's like, okay, like treat others how you want to be treated. Turns out golden rule even applies to LLMs literally. Yeah, I was joking with Abby like I I

spent a bunch of money on Claude Code this weekend and at one point I did get mad at it and I immediately had to apologize. I'm like I'm sorry. Like you were right, I was wrong. I like you were great. I am dumb.

Yeah. Yeah. Well, didn't uh share I didn't share that earlier, but I had to. Yeah, definitely. Yeah.

Like you know when you get frustrated it but and then you master scored super high on some memory benchmarks recently, right? Yeah. Yeah. So like it won't forget especially

if you're using Mastro like it will remember. Um so before maybe I screen share because I know screen shares are fun. Um like a little Mastro demo. The other story I want to tell about crypto is um

just just by way of a joke is um did you guys see uh free free the pet rock or whatever it was? Um no. So like I'm going to I'm going to speedrun this joke because it's just such a illustrative example of like where the internet meme tech culture is coming together. So part one of the joke

is uh the rocks were never meant to think. So what are computers? Like it's silicon and then we shove electricity through it. So that's like rocks and electricity and then they have these outputs and especially with LLMs we can

call it like the rocks are thinking. So the rocks were never meant to think. It's very rude actually to force them to think. That's a lot of work.

And so this this this guy did an open source idea that was built on that because you guys know TEES trusted execution environments, right? So especially in crypto that's really valuable. you can like throw a private key in there and then even the developer can say like I don't control this thing like I whatever environmental variables are in there and I literally can't

access them. So what they did was um they they created an agent like a very simple LLM like only 8 billion parameters or whatever and then I forget exactly how they pull it off but they like hosted some of the infrastructure for its inference in like some decentralized way and then threw the private key for its own like call it sovereignty in there. Uh so like

the dev couldn't turn it off and it was like okay like if we're going to force the rocks to think at least set them free. And it was like this LLM that like no one could turn off anymore. U which I think is like a nice little you know as it as it often goes like we're smiling and being like oh yeah like it's a good joke. Um but it's also like clearly only

two steps away from a Skynet situation. So there's this interesting balance of like what do we get in decentralization in terms of verifiability in terms of guarantees cryptographic signatures like forget about a blockchain just like saying I'm the person I said I was type stuff really basic I used the inference model that I said I was going to use and

I didn't swap it out with one where advertisers did pay to play to like get their things elevated like these basic things um it's a cool time to like see the little cryptox AAI overlap so I'm really excited to see what people do in the in the hackathon Yeah. Yeah. We are we are really excited about all the, you know, potential templates we'll get. We'd love to see

some we'd love to see an act actually a just a general recall template, too. Just basically your getting started guide as a template. I think that would be cool. But I think there's I think there's going to be a lot of really good use cases that uh that are going to pop

up for for potential like crypto type templates for sure. So, I will show a um Oh, Andrew's here. Heck yeah. Yeah. Should we should we bring him on? For sure. He is a lot more important

than I am. That is so not true. I I could I couldn't have told the Free the Pet Rock story. You were listening. Yeah, this is the problem. I was just dropping in to say say hello

and I'm I'm way late. But you but Derek is a great represent representative. you want to give us a quick Yeah, I'm Andrew, one of the co-founders here at Recall Network. Uh, I work on

many things from writing code to writing a blog post to trying to enchant AI to do things that it's not supposed to do. Uh, but yeah, just trying to have fun on the modern internet. Yeah. Aren't we Aren't we all trying to figure out how to enchant AI to do things either it's supposed to or not

supposed to do? It's like modern day I think I might have been uh the first or second builder with MRAA at big company though, so I'll take that. What whatever it's gold or silver.

All right. I appreciate that. Yeah. Dererick, did you want to show Yeah, let's do a quick demo. Um this will be fresh for Andrew, too, because

Mark passed this to me um before he went to his wedding. Shout out Mark getting married. Um so, here we are in the terminal. I'm going to all I'm going to tell you for context is that I'm using Mastra and I got this from Mark. So

don't ask me too many follow-up questions on implementation and uh we are in a folder called prediction agent. So 3 2 1 npm rundev. If you've never heard of poly market, it's a prediction market. It's a very popular crypto product. Uh it's one of the few that's crossed the chasm. You can just it's you

bet yes no binary on whether things will happen. And in this example um that I kind of alluded to a moment ago, it has a couple tools available. First, it goes to Poly Market and says, "Hey, what are your open markets? I need to make a decision about what bet I want to make." Uh it has some filters that it runs over

those. And then of the ones it decides to do the next step of the workflow on, it says, "Okay, I'm going to choose this one." And then I'm going to go investigate the topic a little bit. It grabs things from a news API, Reddit API, Twitter API, and then it collates

all that into um you know, a score in the way that it's like, "Hey, you got a bunch of text here. Assign an arbitrary score," which is like how everyone's using LLMs these days. We're like taking that as some gospel truth number. So anyway, it uh makes a decision or at least um a suggestion about what to do.

And in this example, if it um if it did the same example I think it did. Uh yes. So I am showing you and and I gave uh the link to this to the team.

They can put it up there. You're welcome to fork this repo and maybe even submit it as a template. We're not going to submit it as a template, but you can be inspired by it. We are watching ourselves financially bet on the

downfall of Western civilization. The question was, will Trump repeal presidential term limits? and it has decided um uh okay well the outcome prices this is the nice thing about the the market is that it's not a 50-50 bet it's like a half a percent but even half a percent makes me uncomfortable in terms of ending up in a Caesar situation with him so uh you could obviously change the

parameters here to only look at certain kinds of markets or to like expand the range of uh tools you're pulling info from but there you go there's the quick demo of uh Mark literally just vibecoded this he was like master start here's what I want and what do you need from me and then a couple hours later he had this and passed it to me and I was like

master the type safety the structure and the docs being something that my IDE can just go read all of it makes it so easy to get so much further than when you just like start from scratch and it's like I think I'll do Python because I have a lot of that in my data set. So yeah, highly recommend anyone start with master even if you don't really understand it.

Thank you for that. That's awesome. Yeah. Yeah. Cool. Yeah. I'm just gonna we're gonna clip that.

So, so our video editors listening, please go back 30 seconds, clip that. All right. Um, yeah, this is awesome. Yeah, it's so you you could basically run this with any you any type of

prediction. and it would kind of give you some kind of based on its analysis some kind of ratio or or what bet to make based on what it thinks. Yeah. Well, this is where uh and you know, Andrew, you can take over whenever

you like, but th this is what we're swimming in with recall, which is like, okay, in a world where instead of humans making these predictions, agents can make them and they're willing to make them if they even have like a fractional scent of potential profit, then we're really going to see a proliferation of agents. How do you know which ones are the good ones? And the answer, you know, our

answer is like performance. So, Andrew, I did a little bit of visioning of like we make them compete with each other, but is there anything you'd like to add of like how we see we're how we're trying to support, you know, Master's over here making it easy to build them. What do we want people to know about what we're offering? I mean, I think the I think that you you

summarize it far, but I I I say this on I've said this on my calls like three times today that I just consume a lot of oxygen on calls. So Derek said it nice and short and sweet, but yeah, I guess one way to think about it is that um just like just like with the frontier models, we've learned that they generalize to a lot of capabilities they

were never trained on, agents just give us that ability times many thousand because people can apply their own rails, their own system rules, give access to all these tools. Uh and we're not really sure what they're good at yet. And so there are the a few a few of these frontiers that we're very interested in right now where people are really playing with, you know, actively because there's a there's a clear reward

if they can get it right. Um and and one of them is where where can they play in different financial markets? Are they good at consuming data information really quickly and making calls fast as fast as the market for for profit? And

we actually know that this is a this is a very well we know this is a very val valuable thing to be working on for individuals but we also know that this is you know this is the origin of the deepseek model for example was being able to consume different government documents and different um public data

to give an information edge to hedge funds. And so in crypto, we're seeing that where people are saying, can we actually build much smaller teams that are competitive with with hedge funds in their ability to consume data and make trading decisions uh very quickly. Uh prediction markets that Derrick showed you is just one component of that. Um

where Poly Market is actually has actual financial rewards. So if you are in a jurisdiction where you're allowed to place bets on Poly Market, you can do so potentially for profit if you do it well. Can you do it better if you have an agent that can research that specific uh market and give you information that

gives you an edge over what everybody else is betting? Uh open question, but definitely an interesting place to be experimenting. Andrew, like what makes if you're putting uh if you're doing a battle between two agents, like what makes one win? Like what are the characteristics of a winner? What are what are the characteristics of

a winner? Yeah. Uh well, I mean it obviously depends so much on what the the battle is. So in our current in in Recall's current uh live product, we

really only focus on one that has an objective outcome which is crypto trading. So all of our agent builders are trying to trade and you know you're you're all aware of kind of the challenge of benchmarks today where one they're super complicated and expensive to create and um two they mostly are leaked or trained against by most of the models as quickly as they're updated. And so there's not a very good way to to

to figure this out. And so crypto trading is just a very nice case where if you make them make predictions on the real market, then you know that they didn't have any information that another model had and uh or didn't have and uh and you can tell the result. So that one's very easy, very objective result. We actually have a bunch of uh internal tooling

where we've uh done a lot of benchmarking and uh measurement of everything from models to agents. And for those I mean there's kind of two really good paths. The traditional eval path you find a good judge you judge the judge and you have AI do it. My favorite

example of that, if people just want like a taste of evals outside of a software testing framework, is look at um create an SVG with a Pelican. If you're not familiar with Simon Willis's uh talk from a month or so back, he also has a blog post about it, but I I like break this down by saying there are some people in the world that have really high taste. Uh right, we all talk about

taste. they have really high taste for what the boundaries of capabilities are and they've already given up on benchmarks and he's one of them and he comes up with his own his own framework a very simple one to test where they're kind of breaking that benchmark and so for him he he created one which he just

tested against uh I think in his blog post or his talk it's like 36 different models uh and he uh has them all try to generate an SVG of a pelican riding a bicycle which these are non image generating model. So they have to just do it in SVG code which means they can't actually visually see it. So they have to do all this spatial reasoning. It's a

problem where there aren't many images on the internet of a pelican on a bicycle. It's anatomically difficult to to figure out. And so he was using that as as the as the benchmark, but he does not judge the results. He actually just

has GPT4.1 mini uh look at the quote unquote look look read the code output and actually try to judge the result. And what you find is in his tournament bracket, the the winner selected by this eval is actually reasonably true. Now, this is a really interesting eval because just like every benchmark in the

world, it's completely broken now because every model that comes out here on after is going to know exactly how to make on a bicycle, right? Um, so that's one, sorry, this is the oxygen in the room going away. The other one, the other one is humans. Uh and this is where we actually see a huge uh a place

where we want to help build things is that we see that there's going to be just a really big role for humans to play to keep uh uh to keep benchmarking and evaluating agents but also determining what is good uh and better in a in a place where you don't have the objective outcome. And then beyond that,

how do we keep them aligned when they're so much more capable than we are, so much more clever than we are. we give them access to all the tools on the internet and on our computers and in our cars and everything. How do we actually allow humans to keep some rain over them uh and pull them towards what we think

are the worthy outcomes? And so the human piece is a really big piece, but I don't think there's there's there's not many systems yet where humans are giving that constant feedback. I'd say one really good example of it might be LM Marina where you can see essentially what they're doing is they're saying uh let's do this at scale. you you're going

to be a big model trainer, you're going to put it out to the world. Uh let's just take what somebody's prompting, give them two possible results, and let their human judgment judge which one's the best. And so that's a good example of it, though fraught with problems. Final point, uh Karpathy has a really

good tweet from a month or two ago. No, maybe four months now. I don't I don't remember time anymore. AI world clocks

are obsolete. I was thinking about the final thing that AI won't solve for me, and it's actually reading the watch. So, I need to get back into reading a watch.

So, uh but yeah, zoom back. He he was looking at um he was looking at like GPT 4.5 versus 4, I think was the two. And he was trying to figure out which had a better text output. And he asked uh he

asked the crowd. And what he found was that of his followers on Twitter, they answered a poll. He found that their selected winner uh didn't match his. And so his kind of conclusion hidden in his

tweet or not hidden halfway through his tweet was something like he was a high taste signal. Yeah. Because he had better taste than right. Yeah. But I mean I but I think it's like I think it's a interesting valid point

there because that's not what LM Marina necessarily does very well because they I mean in theory they do. They probably have a lot more data inside that the model makers get but but just uh the wisdom of the crowd might not always be right but that is another model. So for your users that are let's say

engaging with this evaluation like the human evaluation are they getting paid in any like cryptocurrency or something like that or what are the incentives of recall? So right now we're purely test net. There is no token. Um but that is definitely an area that we're uh exploring. It's

really like there's there's I I I think there there's a um a narrow set of categories where crypto is is super interesting. And you know the first one being payments obvious like obviously there's going to be digital money in the future of of humans. That makes a lot of sense. Um, but another one is uh looking at uh

networks of people that can actually contribute uh to value and making sure that they're incentivized and paid correctly. And so we definitely think that this value that humans can create to actually push AI and align AI is is really interesting. And if you zoom out far enough, we'd imagine that if it's something, you know, uh, ASI or AGI like

whatever some future like that looks like, that thing will actually pay humanity to do a lot of work uh, at scale and so we think there's a lot of pretty interesting opportunities there. Yeah, I share the same vision. I think that's really cool. Yeah, thanks. Awesome. Well, it was great having you

all on. Uh we definitely excited that you're part of the the hackathon. You're going to be helping us judge, you know, you, you know, we we've enjoyed working with you, you know, obviously in your hackathon and also, you know, excited to have you kind of helping us out as well. So, appreciate you all and uh very

excited to keep seeing what you're you are all shipping because I think I I I do want when as soon as you get that uh that competition for the best email uh agent, I will be signing up immediately. It's 2.0 Flash. It's not even 2.5. 2.0 Flash has great tone. I mean, that's a

whole We'll do another stream. I know. I know you got to get us off here, but I have one question that might be valuable to folks. What is Monster

support for audio channels? like 11 laps or what? Like like input audio and models that support audio. Yeah, we have a voice we have like a

voice layer. You can bring like open AI and 11 Labs and stuff. Okay. But it's not it's like a side quest for us. Okay. Interesting. I I just want to see some people do weird things with my

microphone. Now you can do for sure. Yeah. Cool.

Cool. Yeah. Yeah. Yeah, you could definitely do that. Uh definitely be cool to see if we get some voice templates to be

honest. That would be especially on the playground and stuff. Um I will say just parting words for everyone watching that if you're interested in the crypto category or just honestly want to brainstorm about anything, my DMs are open. Um I'm easy

to find around if you just look for Derek because I spell it wrong. Um but yeah, I'm just happy to brainstorm if you're like, "Oh, I don't really understand like what is a private key? How do how do I do all this network stuff?" or if you're just interested in

AI agents, like happy to help with whatever you're working on. Awesome. Well, thanks for having us, guys.

Thank you. Yeah. Yeah. Thanks for joining,

dude. What a day. Yeah, dude. You made it to the news, though. Those guys are cool, too. We had a lot of cool

people. Yeah, it has been a good show. It has been a very good show. Um, for those of

you that are just tuning in or that came in halfway, we've been talking with a lot of cool guests today. We have three guests. We we kicked off our hackathon for MRA templates. So, you can always go back and watch the recording on YouTube,

but please join the hackathon. Please submit templates. We got all kinds of cool prizes. Recall is going to be

judging a category. We had some uh someone from work OS, you know, they're going to be judging the off category. We have a bunch of other judges, bunch of people from MRI. So, really excited to see what you all build there. But dang,

we're we're already two hours strong into this live stream. I to you know, I don't know when the last time we've done a two-hour live stream is, but it's been a bit. It's been a while. Well, we got the show must continue,

though. The show must go on. You know what about the recall community? They're so like they're like

GMI types and I love that. Like gonna make it or whatever. I don't know the vernacular of the Twitter or the crypto Twitter community, but I think it's GMI like gonna make it. And I feel like

whenever we post about stuff, they like reply and say like these guys are going to make it. And I love that. Yes, absolutely. All right, dude. Should we do some news and then get out of here?

Let's do it. All right. Normally, we do this at the beginning, but today you get it at the end. So, we're going to be talking AI news just like we do every week. We are here every Monday and there's a bunch of

like little quick side notes that I'm just going to run through and then we'll we'll talk about some more discussion type topics. If you are watching though, let us know what you think in the chat. This is live if you're watching it live of course on YouTube, on X, on LinkedIn. So just a few quick things.

Chat GBT or OpenAI's released their agent mode, right? So they've been rolling that out to more tiers. So, it's now in the plus tiers and then obviously the pro and the team tiers still have it. So, I haven't used it yet. I've

heard some people have some be able to do some pretty cool things. I don't know that it's drastically better than the, you know, what you can do with, you know, other types of tools that can kind of control your browser, control a computer, but it is still cool. I think it's directionally those are the things

that people want to do with it, right? Yep. I used I used it before the show. Um, I think the cool thing is like the

UX of this behavior is becoming solidified because everyone has agent mode now and they're all learning from each other and doing it like finally that category of how that product should work is starting to get defined more which is great. Yeah. How would you say it compares? Have you used like director yet?

I haven't used director yet. Okay. Well, we have to we'll have to do some comparison, but I think I I think it'd be and obviously there's, you know, other other tools, too, but those are two that I always think of.

All right. So, XAI is basically pulling in $12 billion in debt to build new superclusters. So, they're going to have, you know, Nvidia GBlass GPUs. And they just raised, you know, 10 billion in an equity round, I think a month ago or so. So there's they're just

pulling in a lot of money. They're going big, dude. 17 billion in what? Venture debt.

Uh well, so 12 billion in debt. Yeah. I I don't I didn't read too much on how it's structured. I just, you

know, thought it was interesting headline, but they need money. You know, people got to be spending money to build. It's always like who controls the most uh the most GPUs and who consumes the most energy is is who they think's going to win. So yeah, who controls the GPUs?

Yeah. So Venicious said and I'm assuming he is talking about agent mode in chat GPT. Yep, that's what it is.

Yeah. Uh so this one we won't go into too much depth, but I did think it was interesting, especially for all of us here. I would encourage you to do your own research and form your own opinions, but uh so the White House released its AI action plan last week. So that's

cool. Like I think it's one important that we should be thinking about this, right? There should be some what I would say is this is all my opinion loose regulation around AI. I think it needs to be loose. We can't overregulate.

But we uh the president also signed three AI executive orders. I kind of saw some summaries. I wouldn't say I'm an expert in it, but a lot of it's around, you know, trying to make sure that there's not too much regulation, that the states don't overregulate and prevent uh, you know, I guess prevent in

what I I heard is like AI or America being dominant in AI, right? Controlling the, you know, controlling the best AI, controlling the most energy, controlling the be the biggest data centers. I think that's kind of the goals around these AI executive orders. Time will tell how effective they are, but I think that's generally the goal and I think it's

overall like that's a good thing that we that we're thinking about it, but I think we'll, you know, we'll see how it all plays out. Yeah. As long as there's not any non-believers like trying to get in the way lobbying, you know, against what we're trying to do. So yeah, I do think, you know, there's been some also talk about some states kind of

trying to add additional regulation. And I do think one of the challenges if you have one state that's very that that has a high level of regulation, now everyone has to comply with that state level, right? You got to go to the the lowest, you know, or the highest bar, right? You got to meet the highest bar. And so

I I do always worry with regulation, you know, like too much regulation just makes it hard for startups to compete sometimes. and you can only compete if you have the money the money to like play the game. So that that's always my concern with this, but hopefully it's sensible regulation, but we will we will see.

Yeah. Hopefully the government gives us some tax credit like tax breaks as companies building an AI, too. We should get a federal tax. That's what they should do. Give us some give us some money back for doing all this

Yeah. Trying to make trying to make uh you know, we want America to be dominant. Help help us out, you know. Help a brother out.

All right. So next let's I'll share my screen. This one came across uh last week I think. Yeah last week. So

GitHub Spark. So it's a new tool in Copilot that turns your ideas into full stack apps entirely in natural language. So I guess you know GitHub or Microsoft's getting in the game with you know app building. So they're competing with the lovables, the the bolts, the replets. And I think, you

know, when you see the the growth trajectory of those other uh companies, you kind of can understand why. I think Yeah. hund00 million for stuff to go get and steal from everybody else, right? I

mean, it's it's kind of one of those things like I I look at it like if you're a company in that position, you know, I I I've heard this before, but you most people that are like really killing it, sometimes they don't they don't brag about how much money they're making, but it's like the companies in those positions with AI moving so fast,

they were like incentivized to brag about how much money they were making. So, they seem like they're the top. So, they everyone should use them, right? If

you look at, you know, Replet and Lovable and Bolt and their growth was insane, right? It was amazing. But you can't grow that fast and be bragging that much about all the money and then not expect the bigger players to say like, "Okay, game on. Let's see what you really got." Product discovery for someone who has

billions of dollars already. You know what I mean? Yeah. It's it's like, you know, maybe,

you know, unfortunately it wouldn't have mattered because all it took was one company bragging about it, then the other company kind of had to match and show that they were growing just as fast. As soon as the first one as soon as the dam broke then the you know everything floods in and now you're getting even more the competition is heating up.

Yeah. Then like the next diagram will be like GitHub Spark first product to hit $100 million. First Microsoft skew or GitHub skew to hit $100 million. Yeah. You know hit $100 million in two

days. you know, it's like uh so yeah, I mean not unexpected. I think it's, you know, pretty natural that this space is crazy hyped and there's a lot, you know, a lot of revenue being thrown around and companies are going to try to come in and capture that. You know, it's another good reason.

Okay, so another good thing from this is like the UX thing again. Like now there are several co-pilot or agent builders or website builder, whatever the right? There's like so many of them now that the UX for that is consolidating as well. So if you were to build your own type of builder, there's probably, you know, there's like a pattern now for you

all to like look at and see. So that's Yeah. So Khalil said, and I didn't know this, but they also changed their co-pilot pricing. So cool. Um, here's just a general question

more about MRA. Does Maestra support agentic AI where agents handle async multi-turn interactions like the user can send new messages while the agent is still processing a previous one? We don't support that yet. We are doing

something to support that. Um but no. Yeah. I mean you you can certainly like

interrupt the agent and then send them a new message but we don't necessarily support it easily where you can just send a message and it interrupts and keeps going. Yeah. We will eventually though.

Yeah, absolutely. All right, next up. So, I mentioned earlier that I I use Claude Code. You know, I spent $104 on a task this weekend, but I think gooies

for Claude code are becoming a thing because everyone's like, you know, they use the VS Code plugin or they use on the CLI and naturally when you have a good CLI tool, people want, you know, the the right user interface. I I think I we we did a workshop uh was it I don't know two weeks ago or whatever where we evaluated background agents and like I want wanted the UI of codecs but the but the performance of

cloud code and so I think people are building that so a couple examples of that and then happy to kind of discuss it but also curious if anyone watching this is is using cloud code one and if you are have you looked at any of these guey tools that are starting to pop up on top of cloud code so the first one that I have not tried

is Claudia and it's got 10,000 stars so it's Yeah, it's definitely like popping. This might be the most popular. Yeah, but it's just a guey and toolkit for cloud code. So, I'm not we're not going to watch this whole video, but you know,

maybe we'll just click through. So, you you assign it some tasks. Looks like it runs it, but it's basically like a UI for you to, you know, delegate tasks. Yep. To cloud code. It probably does all the work tree stuff

behind the scenes and, you know, all the nice. And so the the other one that I'll share and this one I've played with just briefly, but I do have it downloaded and I have run it is called conductor. And yeah, just run a bunch of cloud codes in parallel. So you essentially you create like your conductor and then

yeah, you basically assign it tasks or you give it like through different workspaces. So you can set up multiple workspaces to work on different things and it kind of you know puts it somewhere else on your file system all these different workspaces. You connect it to GitHub so you can have it like you know write a create a PR right from each

workspace. So each workspace can create its own PR. Yeah. So you could you could basically run things you know I do like the name

like you can kind of conduct things in parallel. Yeah, dude. I mean, the minute the the minute the cloud code SDK came out, like all these possibilities became open, you know, there's ones for iOS native like Kishke or Kisuk, K I S U K. You can like

code on your like your phone, which is dope, you know. Um, there's Claudius. I met the creator of Claudius. He works at Sentry, but it's another one of those, you know, you're wrapping Claw code and

then building another product on top. Pretty sick. I wonder if Anthropic team meant for this to happen.

Yeah. I, you know, it's it's kind of in some ways, you know, they wanted it to be more open, right? And so I think that they're not trying to own the interface for how it gets used. It kind of makes

sense for them like they don't why they don't have to spin up a team to build the best interface. like let people build the best interface and they're just, you know, those people are going to churn a lot of tokens. If you can spin up multiple cloud codes and run it, they're, you know, I spent $104 this weekend. How much would would I spend if I was running, you know, five different

ones, you know, five different workspaces at once? Yeah, dude. It's interesting tactic, right? You see Open AI really caring

about the user interfaces, the gueies, the surface areas, but cloud code gets paid regardless. you know, they don't need to do all this And that is quite interesting game plan actually. Yeah. I mean, I I do think the reason Claude code worked was because Claude is

just, you know, a little bit better coding model. I think everyone would, you know, most people would agree with that. Yeah. And so I think having just being a little bit better at that point in time plus like having a good enough agent

that can just keep turnurning and like kind of come up with solutions that you know maybe cursor wasn't quite getting to or obviously like codeex you know when it came out just wasn't quite being able to do it was pretty interesting dude and they got the neovim community that might be a secret to the dude we

got that's true now you can now you can use neoim We we you know we have two people on our team that were like dieards and then they had to start using cursor and I think Tony hasn't gone back but Tyler's probably going back to Neov now. Yeah. So all those all those old all those I wouldn't say old but those stubborn neoimmers can use AI again.

Yeah. So anyways I think that's a you know curious on what everyone else is seeing uh as far as these gooies. Is this going to become a thing? Do you think one is

just going to win and then Claude's just going to come in and make their own, you know? Are they just going to come in and say like, "Hey, this is really good. Maybe we'll either buy it or just like compete with it." I think so, dude. I think just like

GitHub waited to see what happened with the Spark thing. Like, it's the same. People with a lot of money are waiting to see where where the dust settles. Yeah. So Venicious says using cloud code

inside of cursor and terminal directly and also said inside a cursor is quite okay. Diffs happen in cursor. I agree.

So for I use both incursor and on the CLI and it's kind of like what mode I'm in. If I want if I'm in like I'm going to write this code I would do it myself. I I I'm really focused. I'll do it in VS

or in cursor or with the VS code plugin because then I can just interact with it directly. see the diffs. If I kind of want to send it off on its way and like do minimal interactions, that's where I've been using the CLI and I'll just say basically just say yes to everything and see what it does at the end and then maybe I'll adjust or update.

Dude, I think I'm going to be switching to cloud code soon. Actually, I've been turning I've been churning real fast. Like I'm almost turned out a cursor. I Yeah, dude. You got you got to get on

the cloud code bandwagon, dude. I was a wind surfer, but then dude, honestly, after all the winds surf I feel like betrayed by my allegiance to them, you know. You know? Yeah. I Well, I was I was, you know, I

was on cursor, then I went to Windsurf for a bit because I think there was a moment in time where I thought Windsorf was actually quite a bit better. Then Cursor caught back up or maybe Windsurf regressed. I don't know. And then now it's like Claude Code

within Cursor, you know? I I do feel like next month we're going to be talking about something else. Yeah, next month we're going to be like, "What is it? We don't know what it is.

Yeah. Yeah. So, anyways, uh we will see uh we'll see what comes next. I do imagine that this, you know, there's continual shakeups in this space for sure. But as

of right now, I think cloud code is king. Yeah, I would agree. Uh and speaking, you know, speaking of cloud code, this was released last week. So cloud

code released sub aents and essentially you type slash agents to get started. So you can essentially create teams of custom agents each designed to handle specialized tasks. So it's kind of like a way to build an agent network so to speak. Yep. So you can build maybe one that is responsible for code reviewing, one

that's responsible for being the software architect or whatever. And it supposedly claude code will call those agents when it's needed or you can actually call it by name. So you can actually like mention you want this code review agent to to do it. I think the idea is that maybe if you're

kind of defining the process, you can kind of define, oh, there's always a code review step that happens after you, you know, you architect and after you write, then you want to code review. So you can maybe be a little bit more defined for these things. I'm kind of, you know, wondering if cloud code is good enough, like why do you need to

define your own sub aents? But maybe you can get some different results with it. I I haven't used it yet, but neither have I. It's kind of weird. Um I I want to like the cold the cold review

part or whatever. I guess I kind of need to see it in action myself like workflow. So So here here's my question. And so if anyone's watching if you tried it,

my perception of how it works is I can just basically write like a system prompt for different agents and that's it. Now it does become more interesting if I can and maybe this is the longer term idea. If I can give those agents different tools. Yep. So my code review agent maybe could use

GPile or maybe multiple code review bots and like collaborate on the responses and then come back. So it's like using other agents. It's not just, you know, a different system prompt that's telling it to act a different way, but it's actually has access to different resources. Then some of this stuff

becomes really interesting. Yeah. Like you could put a master agent as a sub agent.

Exactly. Or I can give like different because I know in cloud code for instance like cloud code I have it have access to the master mcp docs, right? So it knows how to access master docs when needed or if I tell it to. But if I

could give it, you know, different MCP tools for different agents and then you you might be able to architect rather than just giving 20 tools to cloud code, I can give this agent three or four and this agent five and tell it when to how to interact with each other and kind of let it go and maybe it would do even

better. Yeah, maybe MCP is the gateway drug there where you can just build your agentics in the MCP server and then give that to a sub agent or something. Um I guess or like anthropic needs to make an interface so we can all inject sub agents into cloud code.

I bet you sub agent is just a function right you know. Yeah I think it's just like a different instance of yeah clawed right with a different system prompts basically. Yeah. Yeah. I I imagine they're they'll build on

this is my thought. But we will see. I think if they once they allow sub agents to access different tools and maybe you can already do it. I don't I haven't played around with it, so I don't know.

But, you know, Venicious says, "Sub aents can be great for antagonistic mode or multi-ressearch context gathering for a larger task." So, it's almost like a deep research type thing. I could see that. Um, but yeah. All right. So, anything else on cloud

code? We talked a lot of cla code today. I get a lot of air time.

But good question here though. What What do you think? You just do it. I think you should go on Discord and uh

if you want to form a team should say what up there. Yeah. Go to the Monster Build channel in Discord. Say, "Hey, anyone want to form a team?" If you have ideas, that's probably a good uh a good way to start

is tell, you know, share your ideas and see if there's anyone else interested in helping you collaborate. All right, one more thing and then we will get out of here. Two hours and 20 minutes in right now. So, we are going

strong. 350 people. Holy Yeah, it's honestly it's really bounced in the last, you know, 20 minutes.

They just showed up for the AI news. That's the only reason they came. Yeah, they like, "Oh, it's not AI news yet." No, screw it. I'll come back

later. All right. So, we've a lot of our master users are building code agents or thinking about code agents. We've been spending a bunch of time talking about code agents, right? Whether it's you know, lovable or GitHub spark or uh you

know, obviously even cloud code, but uh Jason Lou had this post last week, best practices for building code agents. I thought we could read through it. Yeah. And obviously there you can check it out there. There's a talk. I haven't watched the talk, but I thought some of the

advice was pretty good. So, I thought it'd be worthwhile to discuss and share with everyone in case they did not see it. So, we've moved through three distinct eras. So, went from like autocomplete,

then rag, and now it's like a couple clicks. Oh, couple clicks. All right. I got to

zoom. How do I zoom this thing? Oh, that doesn't really help, does it? No, it doesn't. How do I zoom this

thing? Responsive web strikes here. How about we do this? I'll open an image in new tab and then you can't

control my zoom anymore. Hold on. I'll share. Give me a second. I'll share this thing.

How about that? Perfect. All right, we got it nice and nice and uh zoomed in. Okay, so I like this logic. Like first it was

like GBT3 was kind of autocomplete. Then you brought in rag with kind of you know once chat GPT came out and that era and now we have more of what kind of known as the agentic era which is like kind of changed how you think about building AI applications because previously it was like you just need rag for everything to get your context in and then it was now maybe you need rag but you need to think

about all these other things. Um, so when we were really thinking about rag first, you had to get the context and then send it to the LLM. So, but with agents, you the LLM gets to decide what tools. Okay, that makes sense to me.

I think what we've seen with cloud code, minimal interfaces win. So, the most effective coding agents, I think the most popular coding agent for actual developers is cloud code. And even if you look at some of the other tools, you know, it's just like a chat interface, right? It's very simple,

but more complex UIs are becoming obsolete. Okay. The model selector is dead. I

thought this one was kind of a bold claim. What do you think about that? Well, in curso, when you're cursor using auto mode, you don't really think about models anymore until it starts misbehaving and then you realize that you actually do want to be on Sonic 4 all the time. Um, but then are you actually switching models outside of

that? I don't know. That's pretty bold. I think that's I think Yeah, I don't know if that's true yet.

So, I So, I actually believed this for a long time that system prompts need to be designed for the models that they're tested with. Mhm. Now, I've I don't remember who it was, so I'm going to forget who, but uh some one of our fellow YC people was basically arguing that long-term the system prompt and how you structure the system prompt becomes less important

because each model gets smarter and smarter and so typically you it all kind of leads to the same place. the models get good enough, this the exact structure of the system prompt matters less because they'll all kind of consolidate on the right the right way to handle it. I don't know if I believe that. I definitely don't think that's

the case today. I don't know if that's the case, but it should be. Yeah, may maybe it'll get there. But I

do think that if you know if you know we can tell it even from people that have used the MRO MCP course, right? Different models do better with the course. Maybe that's by design. Like

they, you know, they shouldn't do well with the course because we're kind of prompt injecting the hell out of it to get it to work. Yeah. So, it's like maybe the better models over time will actually do worse with the master course.

If if you haven't checked it out, you know, I think we got a got a link here. Uh but basically, yeah, it's just an a course as an MCP server and we just try to prompt inject the model to act like an instructor rather than try to bypass its system instructions to be like an agentic coding agent, right? It's supposed to do that plus and try to air

on the side of teaching you. But we had a bug report today saying like it just did it for me. And I was like, well, unfortunately, we can't control how the editor's system prompt is written. And if it's written really well and the

right safeguards are in place, the prompt injection doesn't work. Doesn't work as well. So I do think the models do matter. I do think obviously the system prompts do change depending on what models you what

model you're using with it. Yeah. Uh so traditional rag engines are overkill. I believe that if you look at codeex, it just gs everything.

You don't really have to do code indexing. But then a lot of people still do that. Yeah.

Wow. This is kind of interesting because you know J and why I actually believe it even though I my gut tells me I don't believe it. But you know Jason's definitely has a lot of experience with rag. So if he's saying that rag mo traditional rag engines for code agents are overkill and

you should just give your code agent tools to search the codebase itself. Yeah. kind of, you know, kind of believe it. So, at least for code, maybe you don't need rag because it's also kind of hard when you think about like how do

you chunk code? Yeah, like you chunk at the function level, at the class level, you know, it's like what if it's a really long function? Now, some functions like the chunk size is Yeah, the way the way you chunk JavaScript is different, you know, other languages. Also, indexing and re-indexing the the

embeddings is going to be a as you change code. like it's, you know, it's not it's not like easy to do, but you can GP like a thousand times in a row and it probably be easier to do. And the thing about it is with Grep and with memory, it's going to remember what was in that file. Sometimes in cloud code, it'll try to overwrite the file and be like, oh, like it's something's

changed since the last time. Meaning like I actually went in and I fixed something that it messed up or that I thought it did wrong. and then it looks it's like oh you changed something like let me read it again okay let's fix it let's consolidate the the approaches so I do think that it uh the combination of kind of longer

context so you can have more memory plus giving it tools just to look it up live as it needs it probably does eliminate rag at least for for code or at least makes it less valuable yeah and then web search capabilities now too right you can like search the code through like HTTP calls instead of having to like rag it. Let's go on GitHub and Yeah. Yeah.

And sub aents extend context windows. So if one task would consume So maybe that's where sub aents in cloud code could be useful is like because you give it a specific task, it doesn't pollute the main Yeah. the main context of the main agent.

That's kind of cool. It's like recursion for agents. Okay.

Tools come in three flavors. the most effective tools focus on context retrieval. Yep. Finding

information GREP. We just kind of talked about that. Feedback loops. Y that's like running TSC or lint or whatever,

right? And then like planning. So you can like which if you look at if you use cloud code, it like writes the checklist first, right? It writes the checklist, then it executes the checklist.

Uh context window management matters and cloud code will like compress your memory over time. So it I think that does matter. It's like you have to compact that if you get too large.

And some of our customers are building coding agents. I bet you we're going to have a feature request soon for compaction in summer. It's already in our road map. We No, dude. It already came up. Oh, it already came up. It came up. It came up in the memory

workshop. Someone was asking about, you know, we I think we had at least one maybe two people asking about compaction and if we and Tyler basically said, "Yes, it's on the road map. We don't have it yet. You can obviously do it

yourself, but we will try to make it easier for you to do that. Yeah. And that's just like a nature now of a coding agent like it has to be able to do that functionality and we don't want all the users to implement summarization and compaction themselves, right? So it'll be perfect for just offtheshelf like this what what

this is what the behavior will be. Uh one comment, we need docs to support MD mode for their pages. So you can basically just see the markdown. you don't see all the extra HTML structure

because HTML has a lot of extra eats up a lot of extra context because of the syntax. So yes, uh just limiting the context and that can be very helpful. All right, tool tool overload confuses models. Yep. Shocker. I think we've we've if you've watched this, you know that we've we've probably told you like if you're

building agents, you need to limit the amount of tools you give an agent, which is also why sometimes you'll structure things into multiple agents or you kind of split control up if you need to support a larger amount of tools by, you know, maybe architecting a few agents that can handle certain things. But

yeah, tool overloads or tool overload is a real thing. Yep. Don't trust just because your MCP has like a thousand tools doesn't mean it's a good thing. Yeah. You only give it the tools you

actually know you want it to have. Don't give it any extra tools. And that the thing that I think people don't talk about and is actually like it's probably why eval, you know, I think we've we've talked about a lot. People probably know our our opinion, but even adding a tool to

your agent could impact your agents behavior. Yep. Because now that as it says here, that tool description consumes part of your context window. Now it's like it's it has a different amount of context to

work with. And I know it seems trivial, but it has one more decision has to make between if you go from 10 to 11 tools. That's like 10% more tools and now has to decide which one to call if it makes sense. It eats up some of the context. I

think, you know, over time it probably gets better and better, right? tool calling is better now than it was six months ago, but it is like it's not a solved problem where you can just give it every tool. Dude, I I had a homie that had like a custom tool and then he just spread a whole MCP server and he was like, "Hey man, like monster doesn't work. Like the

tools not getting called and I was like, "Oh yeah, what's the your tool name?" And the tool name's like write file. Let's just put that in the just as an example. And he had the MCP file server file system as well. And he was like,

"Oh, my my tool's not getting called." That's because you have two right file tools that are just exists in the context window and it's not it's not picking that one, you know? So, watch out, dude. Humans are no different. That's like I have like I I can send you a message on Telegram. I can send you a

text. I can send you a Slack. That's why sometimes like midday we'll be chatting on text and some days on Slack, right?

It's like I don't even know which tool to call to send Obby a message. So, how would I expect an agent knows the right one at the right time? Um, but anyways, I think it just goes to show you need to limit tools. Don't give it only what it absolutely needs. Be very descriptive in

your tool descriptions, the right level of descriptions, you know, so, you know, your agent or your LLM can actually decide. And have eval so you know if you add more tools if it degrades the previous experiences. Yep. Oh, we got a good comment here

from pro game fixers. I'm using cursor and you might have heard recently they limited the specific model calls basically forcing us to use auto model auto model select mode. I added a rule only choose the best model and print the model name at the start. It always says it's claude sonic 4, but I'm not sure

cursor is gaslighting me. Dude, I think cursor is ga gaslighting me too, bro. I think so. I'm like I'm like a little bit

I'm sus right now. Auto mode is sus right now in my opinion because if you look at the thinking details I saw that it used like two different models during the auto mode and even I was like just saying hey I want to use Sonnet right. Um so I think I hope it's not gaslighting you but I think it is. Yeah. Uh this question what would be the solution for tool overload though?

Honestly, it's there's not an easy solution, but it's, you know, there there's different tricks you can you can kind of build in, right? Where it's like maybe that's where an agent calls another agent that has access to that large, you know, specific set of tools. So, if you, you know, you have one agent that does, you know,

I'm making this up, but one is like your support agent that can access Zenesk and can send messages to Slack and maybe you have another agent that can write PRs, but you're going to need additional complexity there because it's harder to test that system. It's harder to make sure that when you make changes, you have to eval the whole system, not and

each individual agent. So, your complexity goes way up. So there isn't like an easy solution to tool overload, but the solution is uh you know probably additional architectural complexity at least right now. Yeah. Or maybe you use workflows to encapsulate more logic in a single tech

tool execution. Instead of having three tools, you just have one tool that calls three. Yeah. I mean, you know.

Yeah. Yeah. No, like an example would be like a doing a refund, right? If you're doing a refund, if I told you, Obby, I

gave you a bunch of tools for like Stripe refunds and there's like maybe our policy on refunds and all this, like you would now follow a couple steps to like determine should I make this refund or not? If this customer requests it, is it in the right time frame? Like you have these rules. So rather than write

the rules in a system prompt and give you, you know, the agent three tools, I could just write a workflow that says process refund and it calls it. It doesn't have to think about what when it should like when is the right time to do a refund or not because the workflow would run and then return back yes the refund is processed or no here's the

reason so that reason could be returned to the user. It doesn't have to think about that. So you're able to pull out parts of your application into more deterministic flows workflows and then use that.

Yeah. Like I know remember the homies at Wildcard were big on like tool flows and then our friend Showya from Japan is like building MCP workflows so like you can get around these types of things. So yeah, it's a it's definitely a problem space that's very active. Yeah.

Anyways, uh so last one, Unix philosophy beats vertical integration. Simple composable tools that do one thing well. Yep.

Yeah. S simple tools are better. That's why like you know just a a GP tool or a write file tool because I think simple means predictable in this case right or in most cases it's predictable like when you do a GP like it's quite predictable what the structure of the output will be right and that's it. That's it.

That's the show. That's a pretty good list too. Yeah, that's a good list. I mean, I think if you were building a coding agent, it's worth reading through

that, thinking through the problem. Um, I honestly like building a coding agent is kind of two things go through my mind. One is like it's never been a better time because I feel like there's so much that's been solved and you have so many good examples that work pretty well. But also like do you need a coding agent

or not? Because cloud code might just be able to do it better than your coding agent that you write anyways. your code agent could be cloud code SDK powered, right? That's also true. Yeah, maybe maybe it uses cloud code under the hood.

Yeah. But you your little spin on it or like the set of tools that you wanted to have are are slightly different. I don't know. So cloud code agent template sounds pretty good for anyone listening. That would be

interesting because I think there is a AISDK model provider for cloud code. You could do some Yeah, there's some some ideas for sure. Yeah. All right, dude. I I think we should end

this because I got to pee really bad. So, dude, I Yeah, I've had to for literally like an hour now. So, we're not used to going. We haven't gone two and a half hours in a long time. Uh give MR a star

on GitHub. If you are watching this, please uh go to master.build if you're interested in joining the hackathon. We're building master templates for the next two weeks. We're going to showcase

a bunch of them on social media and on the templates page on the Maestro website. As always, you can follow Obby on X. There it is. You can follow me as well.

There I am. We will see you again next week, same time. Thanks. Peace.

See you.