The Goose is Loose! Plus Hackathon Awards, Cedar OS, Security Corner, and AI News

Hey everybody, welcome to AI Agent Hour. Hopefully you can hear me. We kind of uh

rushed the mic setup here today, but we also have Shida, which Shreda, you are no longer muted. How are you doing? I'm pretty good. Uh Shane, as I

understand it, um you just pulled in uh from an Uber just like you're holding your own like late night, you know, show the red carpet right into the right right from the Uber five minutes ago from the airport to here. So that that's delivery delivery. That's a dedication. And also here, sorry, I had to pick up these books.

Obviously, we are you know where are we Abby? We are in the palace of the dog. We're in the dog patch.

We're in the dog patch in San Francisco. So, we're doing it live today. Normally, Abby and I are on, you know, other ends of the not other ends of the world, but other ends of the United States, but today we're together and we got Shida to talk more hackathon and templates. So, I'm excited for that. But Obby, you were

uh you were out last week. Yeah, I was sick last week. I was up in the stratosphere. Um, so they say, but

uh, it was good. Um, it was good. It was actually I caught up in a lot of sleep, so that worked out well.

Um, but a lot of happened in last week, too. So, yeah. Feeling better now, though. Yeah, I'm back to normal. Um, got a new

type of Celsius here, too. Cherry Cola Celsius. So, that's cool.

Cheers. Is this show sponsored by Celsius? One day. ongoing joke is that, you know,

we're looking for a beverage sponsor and, you know, can't get the logo. We got to hold the can, right, Shane? That would be a good start. I'm inexperienced. They don't pay me for

this, you know. We have a lot of episodes, you know, and we drink Celsius in pretty much every single one. Most of them. Yeah. I'm not I'm not opposed to Sugar-free Red Bull either, you know. You know where to find us. Uh, so yeah,

today we're going to be going through a whole bunch of things. So, we got the hackathon, we had the encore week, so we got a whole bunch of additional submissions. We're not going to spend quite as much time as we did last time. We're just going to watch the videos because we got

to get through them. And then we'll give out some awards. We're having uh a company called Cedar come on, two of the founders from Cedar. We're going to have Security Corner with Ally. We of course are going to have AI news. And then the

goose is on the loose. The goose is on the loose today. The goose will be in in the show. So, we

have uh Dan Goosewin from Dev Dev Talk and also he's a he's at Bobby now. So, we're gonna be talking to Dan as well. So, yeah, the goose is on the loose. It's going to be a it's going to be

probably a two-hour show. So, if you are here, uh, buckle up because the shenanigans is just getting started. There's going to be a lot of shenanigans. It's a beautiful day in San Francisco today for once. Um, so that's

also on our side. We usually do these um founders weeks in SF. I mean, I live here full-time now, but you know, Sam's actually in the room over there, so we're all together. It's good because we

get to kind of plan what the road maps should be and then what fixes and emergencies and all that type of stuff that we're we're dealing with. Um, so I'm glad. Yeah, it's going to be a good week. All right, so without further ado, let's

talk about master.build templates. All right. Hopefully we can see this.

So, we had a hackathon. It was originally two weeks and then we extended it a week and got even more submissions. The whole goal is to build templates. We released templates not that long ago. And we've just found that

people typically want to have things to one learn from and two potentially use as a starting spot so they don't have to start from zero when they're building out agents and workflows. So that's what templates help you do. And so we wanted to have a hackathon around that. So we have a whole bunch of prizes that we've

announced. This some of this is review if you've been watching the last few weeks. We had gift cards. We have a

whole bunch of you know prizes just for submitting you know kind of raffle prizes as well. And then of course we have ability to be spot, you know, potentially have your template show up in the template library under the community section. So hopefully we'll show that as well because we have some of them listed there now, assuming my PR

went through. But here's the the page. And so we had a whole bunch of categories. I'm not going to read them

all, but we had a bunch of categories. A bunch of uh judges to help us out from Smithery, from work OS, from Browserbase, from Arcade, from Chroma, from Recall, from Confident AI, the homies, and then of course, yeah, a bunch of people from MRA as well helping out. So, let's do some awards. But before we actually go through the awards, here's the ones we're going to be judging

today. We judged the uh all the rest of them last week. We're going to judge these ones today. So, we're going to

give out the best overall. Going to give out best MCP server as well as a bonus award. Uh, best use of Smithery. They

win the Nintendo Switch, which is pretty sweet. That's sick. Best use of Oth, use of web browsing, rag, and then a crypto agent award. So, not necessarily in that order, but let's

talk through them. So, the first one we are going to do, we're going to show and we're going to show the videos first and then we're going to talk through and actually give out the awards. This one's tight.

All right. So, here's the first video. Hopefully the audio comes through. Seven minutes. Okay. We're not going to watch the whole thing. We'll probably click

through this because it is a and I'm going to present to you today the XRP ledger uh AI agent that I built with uh with Master AI. So that's a template. So I hope uh I hope it's going to be okay. And um yeah, so the XRP ledger is is a blockchain and XRP is the

native currency uh currently number three in market cap at around $3. And let me show you what I built uh with with Mastra. So I built an agent XRP ledger agent and it allows you to um query the data that that that is on the ledger. It allows you to submit transactions such as payments, configure

accounts. You can see I've built many many tools uh here on the right side. So let's say for example that I would like to know the the block number and we call it the ledger sequence on the on the XRP ledger. So get the uh the ledger

sequence on mainet. So it's going to call the the right tool and we're going to have the ledger sequence which ends by 400 0. If I go back to the explorer, you can see that we have now the 4400 0. That was the the last ledger sequence at the time.

We can as well create new wallet. So create a wallet. It's going to call the right tool. We have the public key, private key, the encoded address starting with an R, the

seed which is like an encoding of the private key as well. We can fundry wallet using a faucet on on test net fund this wallet on test net with a lot but there has been another tool. Imagine having this smart contracts cuz like and after 3 4 seconds should have the response. So let's take this I'm inspired new address ending with

Okay, so I would love to watch the whole video. You get the whole idea. This is a sick template. This is going to be on the

templates page eventually because it is awesome and uh it's a really good video. So I'll probably even include the link to the video as well. We don't we don't include all the videos but this one's very detailed. So, it'll be

copy that private key is getting 15x RP, I guess. Yeah, I imagine it's uh there's nothing in it, but you know, you never know. You know, that's I'm I'm sure he took precautions, but that was a awesome awesome on test net. Yeah. All right. So, we got another one.

It's gonna be I mean that was a good one. That was great. So it's it's kind of hard to follow that one up but here we go. This one's at

least one minute so we can get through the whole thing. Here we go. Make sure there's some volume here.

This is my submission for the trading agent for master that helps you and assist. We're using Gecko API. We're using recall and we're using um Gemini for this output. Um it really doesn't

matter which one you use. Um the toolkit that the tool that we're using uh we call the history you get your process on how it happens on MRA. You're able to see your portfolio um see your execution and balances and portfolio actions. It kind of gives you

a good um a a good risk management within the code. And with regards to the workflows, wait, sorry, it's the wrong workflow. The workflow you're able to see, you gather the context, you generate your plan and plan by context and then trading and you validate. So can also

give you some recommended trading. Um haven't yet tested the output in that much time. Uh thank you for my thank you for the submission. Another crypto one.

Yeah, two two in a row. But that's not the category though, right? Uh submitted under MCP and uh and crypto. Oh, nice. So, you could submit under multiple categories. So, that one was Oh,

under multiple. All right, we got another one that was This one was submit under MCP category. Let's share this one.

So this is the template I built for the mastery. It's a smart MC page. I'm going to make this thing bigger. What is this? There's a preview thing on the top right. Maybe that's Oh, Canva. Canvas.

So this is the template I built for the master. It's a smart MCP agent that literally upgrade itself on the fly. So it start with only one MCP server. It's a text MCP server to edit those files of

the project by the agent itself. So when a user ask a question if it doesn't find any MCP servers regarding that it going to query smithy in real time to discover brand new MCP servers and it can install those MCP servers in the project itself based on the user request. So let's see it in the action.

list the available make that go away MCP servers. So when you ask it going to call the tool and it's currently showing there is only one MCP server text MCP server. Okay. Now I can say I want to

learn about recent news. This is of AI. So when I say this, it going to search for the related MCP servers. So it's showing the list of MCP servers. Nice.

The top five. This is cool. Okay. So you can select these MCP server

from the list provided. Otherwise you can uh if you know the MCP server, you're going to ask you can directly ask for that one. Let's say uh I want Google News MCP server. Can you search

about it? So it's searching for the particular NCP server. Yeah, now it's showing the MC available related to the Google news.

Okay. So let's say we can uh you want to install this one. This is a long Yeah. It's

one skill acquisition and that's what this is called sheet here. So what now what we'll do is it will check whether this MC server is already exist in our project otherwise it going to add MCB server new MC server to our project. So now you can see the MCB server it installed itself the Google news one. So you can get this error because the

server restart itself because of adding new MCV server. So we can ignore this one. And we can say or you can ask. Yeah. Why even write it to a file,

you know, recent news. I think that's true. Deep mind. All right. We we'll see. We'll let this thing run and then

we'll stop this thing a little early. Can I say this? It going to use a tool from the installed MCB server. So you can see It's called the Google News Research

News for the MPs or regarding All right, that was cool. Very good. Okay, so let's talk about that just for a second while while we pull up this next one. So skill acquisition has been something you and I have talked about a

lot. So that was pretty cool. That was great. You had one comment. Why, you know,

maybe you It's nice that it can be written to a file, but even better if it can just dynamically learn and then read some kind of config. Yeah, somewhere else. But that's just an implementation. That's really

I don't know, dude. You missed it last week, though. We had some There's been a lot of great great uh submissions here.

Yeah, I saw the one with scoring or scores. That was sick. You all know this is going in the template library, so like people can use it. This is even cooler, dude.

If if you're watching this, you you're going to be able to just download and use these things. I mean you can just you can run it locally, you can run it on master cloud, you can run it anywhere. Khalil said this cool idea. Yeah, dude it is cool. You should like fork it and

then try some other stuff on it. Khalil actually also won an award I think during the first batch. So love to see it. Yeah. And yeah, and we'll talk about this.

It doesn't doesn't end here. If you have more templates that aren't built yet, you know, we don't have necessarily prizes available. Oh, always. But we want to grow the library so others can can learn from this stuff. So, all

right. Austra Austramoth. All right. It's been a long

day. Mostra template. It's a mouthful. Mothra.

Mothra. Okay. Shane woke up at like 4:00 a.m. today, so it's all Well, in I'm in Pacific time now, so it

was 1:20 a.m. when I woke up. So, we're going. Let's go.

Okay. Um, so Bruce, uh, here to kind of demonstrate a little bit about what I built from an off standpoint. So, I'm going to start a quick demo. Uh, so, basically, it's a ticketing system, right? Kind of, um, looking at a help

desk agent. Um, and this is around off. So, does the agent have access to call a tool? um is what I'm trying to demonstrate. So in this case, I'm doing a get ticket and so I'm looking up ticket two, which is just an array right

now of a database that the tool has. And so it's going to and it's getting the ticket token. Uh so it's authorizing the identity of the agent and then it's assigning a token um that lives for a period of time that shows the uh agent in pro in process. So in this case, I was able to go out and get the O token

uh from a server that I've set up and then that off token was able to be used uh to be able to say, do you have access to pull the records from the uh the specific tool? In this case, it's just bringing back all records. In the future, I can kind of get to the only bring back the one record. Um and now

I'm going to simulate the similar thing of create um a new ticket for meeting. soap in my hotel room. Okay. So, similar

kind of approach here. Go out and authenticate uh the ticket. Um so, okay. So, fine.

It's asking for the title uh description or asking for the description. No soap. And um Okay. Okay. So, then going out

and getting a great ticket. um is uh still needs to authenticate itself and so it knew it had the token and because of uh how I've set up the off and the policy which is another aspect of this it was able uh to create the uh token on the back end. So I know I'm out of time but um read me will have more information.

Sick. That was tight. Uh, so thing I like about that is it wasn't overly complicated, but we always have people asking about how and there's more than one way to handle authentication, right? In this case, the agent is actually authenticating, you know, to the service

versus just like having authentication built in as like a key or something. But that's cool. I think people want examples of how different methods for using O. And so that was one of the reasons we had the category. And I

thought this was really good because it was just super straightforward. They're straightforward, simple, but easy to follow. You don't need to worry about OOTH 1.2 and anything like that.

It may not be that secure, though. I'll be honest about that, but it is straightforward. Yeah. Yeah. I mean, yeah. Well, can't

vouch for that. Yeah. All right. So, we have one for web browsing. This This next one was in the

web browsing category. Sick. I bet you this one will be done. All right. Let me share this. This one's

available on YouTube. You can watch it here. Just it's unlisted, but we'll Hello, I'm Yenni and this is my master template which is called cookie compliance checker. Consists of uh five

tools. Four of them are copied from the existing official master template which cover uh the basic stage hand actions for navigating to the web page perform actions and extracting some data and uh also added fifth tool which finds and clicks the accept cookies uh model. Now let's try this out. uh we will provide

some web page and it will what it will do it will open a headless browser as this is a dev mode it opened actually a real browser and it found a cookie uh model and clicked on accept and checked which cookies were set before and after clicking and gave a compliance score here. It out outputed the cookies that were uh set before and after clicking

and also provided some recommendations. I hope this template will help out people who want to start a new master project. Indeed it will, dude.

I mean, I'm not a fan of the cookie headers, but you know, they are we should merge that into the main template. Yeah, they are. Yeah, that that seems like a good example of just like browser use. So maybe that just becomes rather

than it's standalone, we just pull that in to the main Yeah, the main thing because it just seems like a simple really good example of what you might want to use with it. All right, we'll do one more here and maybe two more. Uh yeah, one more then we'll do the awards. That's all we got time for. All right, one more here and then we will give out the awards quickly and then we got to continue on

to the next uh next segment. All right, this one is a rag rag template 3 minutes. Hi everyone. So this is my submission for the rag template. uh so what I have

done so the problem we are trying to solve here is there are always uh ACL you know access control list is always in place in the companies but when it comes to the vector we just put all the data talks into the DB and we get back and anybody can query those right but there is no ACL within the vector space so this is what I'm trying to do here that uh for example if you go here if you see we have I have created

like few sample documents into my repository and uh certain rows Now I'll show you there is no template like this. Now I will go and start indexing. We have two workflows. One for indexing and one for actually quering the uh stuff. So if you see the it has been indexed and if I go

back here and you can see we have this indexed. Now I go here. Now we have a two different kind of roles here. One is HR who has the maximum uh access and

others doesn't have. So if I go here and if I try this, it will run the workflow and you will see that it it gets the result back for its query because this is the document which is available in the employee handbook. So the uh oh sorry the financial handbook. So it it can access those document. Just a second let's see

it's generating the answer and soon we will see verifying the answer. So all that in place and you can see it's available. Now if I ask this question which is available in the HR handbook, it will not it will fail.

All right, I think we get the idea. That was cool. Oh ACL's. Yeah. Basically, so it's not just rag,

it's rag with ACL. Just like the OSO Yeah. concept that we saw. Yeah, we chatted

with OSO what two weeks ago and it was similar types of ideas and it's one of those things that you know you you start to design these systems and you don't think through like okay you got to figure out how you handle security how how you handle off how you handle who can access what so I think it's a good example just as a way to yeah you know understand who can access

what what documents or what you know chunks I guess in rag All right, let's go back to the awards. And, you know, we kind of buried the lead a little bit here, you know, or we we didn't, you know, we we kind of maybe had a little bit of a tell. Our poker face was not very good. So, the best MCP server was the trading system

that we saw that used recall. So, congratulations on being named the best MCP server. I guess that was the first one we we watched, right? The XRP, right? Yeah. The the second one was the best use of Smithery template for agents that can

dynamically discover, install, and use MCP. So, that one was really cool. So, you you win a Nintendo Switch. Congrats. And we I guess we got to say who maybe we should say the name of the

person actually was um a Aka I think who who won this one. Um this the next one who won the switch is I believe it's Shinas. Uh he won um the a best agent network uh use case as well last week. So he's on a roll. Yeah.

All right. Making moves. And you can see the GitHubs are at the bottom. And you know we will have all

you know these will all make it to the templates page eventually. Uh the best use of O. We only showed one, you know, the agent authentication system and didn't what what was this person's name? They won last week, too.

Yeah. Bruce also did um the internet of things implementation last week as well. So, we got some we got some hard hitters. Yeah. Some of these people are cleaning up the uh the best use of web browsing was the

cookie compliance checker. That that one is good enough. As Obby said, we should just maybe add that to our example because it's a really good, you know, basic use case that I think everyone can kind of relate to. So, who who did this

one? I I'm gonna butcher the pronunciation, but it's I think it's Yev Yev Jenny, something like that. Um, so sorry, but congrats. Also, congratulations.

All right, best rag was the governed rag templates with the ACL's. Cool. So, and did this person win last week as well? I Or maybe he did. No, he did. He did. He did. We

have so many re They're making templates. They're turning them out every week. Yeah. I mean, keep congrats. The best crypto agent was the XRP first

one that we saw. So, that one was the seven minute one. That video was great. We didn't get to watch the whole thing, but who who was the

And uh you know, Shane, I know we were discussing how to do future prizes to international um contributors. So, I'm like, maybe this is the way. Just make it easy.

Yeah. Yeah. Maybe we Exactly. Yeah. We We One of the things we've struggled with is how do we get prizes to people?

And so, we're like we just need to send them stable coins and Yeah. or Yeah. So that we'll figure that out, but I think that that might be the way. All right. And the best overall, this one's from last week. So there's this AI

storyboard generator. You can go back last week and watch the video. If you dig into the code, it's just pretty sick. All the complexity. I mean, I think there's like a dozen tools, two different workflows, four or five

different agents. So it's a really complex and really well thought out. And this one just, you know, we all kind of talked about it, but this one hit home with me because I've tried to solve this problem myself and getting consistent characters across generations. It mixes

in agents for like creating a story as well as image generation models. So, it's just a really well done template that if you want to start with like a storyboard creator or storybook creator, it's a really good starting spot, dude. It's also it's just like Slack notifications using Zapier like Yeah. exports to PDF. I mean, it it writes a PDF file. I mean, yeah, there's like a it in integrates with a whole

bunch of different services and it's just pretty complex and well done. I would say it's really good teaching uh if you're trying to learn it. It does everything you want to like get adept at in a really fun way, you know.

Yeah. Yeah. Exactly. This is this is Trina winning his third prize. He really cleaned up.

Cleaned up, dude. All right. Well, yeah. So, as we said, there was awarded

prizes. There's a whole bunch of raffles that you that we don't have time to announce today, but check your email. We will send them to you. There's a whole bunch of raffle prizes. Yeah, Obby, you can hold that up. Everyone get that

submitted gets one of these. Uh there's a whole bunch of other prizes that you know, Shita's gonna be emailing in your inbox. So, even if you didn't win now, you might still win one of these cool prizes. Shita, what were some of the other prizes? Oh my gosh. I think Scrimbo was giving

out like 10 free memberships. Um I think we have the AI engineering, you know, gospel bible book that we're that we're giving away. A mechanical keyboard, which honestly huge level up. Uh some

really good stuff. Oh, and a Raspberry Pi. Like lots of good stuff out there. Yeah. So keep your, you know, if you did

submit and you didn't win anything yet, you might still uh you might still win something. So, you know, check your email. Oh, here they are. Here they are.

And so, you have the names. Okay. Yeah, I did. I did put them in. Yeah. I'll email them. We don't We don't need

to read all of these. I Yeah, but if you're watching and you see your name here, you won. Congrats. Congratulations. Thanks for familiar names, too. Shita, you must you put the slide

together between when I got in the Uber and when I got here. Yeah, I did. I did. Nice.

We do things professionally around here. We're very well prepared. We do it well in advance. That's how we roll. Yeah, we're so prepared and

professional. All right. Uh, but we couldn't have done this without our judges, which we talked about earlier, as well as our sponsors. So, want to thank Recall, you know, thank Smithery, thank Scrimba. We

couldn't have done this, you know, the last three weeks without them. So, really appreciate all that they've done. If you haven't checked them out, please do. You know, if you're into crypto, you want to run like competitions for

agents, recall is really cool. Smithery, if you're using MCP, you should be you should know what Smithery is. And then, yeah, if you want to just level up your skills, check out Scribba. Check it out.

All right. And then, you know, just our stuff. We don't need to talk about that anymore. We did it. We got through the

in good time. We did it. We did it. Not Not bad. We're only six minutes behind schedule. That's not too bad.

Doing pretty good. All right, I'll bounce. Good seeing y'all. Um people, please continue

submitting templates. Like we'd love to feature them in our communities library. Yeah, if we and you will see it will be live soon that we will have community uh community templates from last week and this week up on the Maestra templates page. So that should be live this afternoon, you know, in the

next couple hours and we'll have some of them live and we'll be rolling out a lot of them. Even if you didn't win an award, your template might still be featured. So we're taking we're we do have pretty high bar, but we're trying to take as many as we can to get them listed on the on the templates page. All right, see you guys.

See you. All right. All right, dude. Man, those templates

It's a game changer, I think. Yeah. We need to do more hypoth. Yeah. Yeah. If you're in the chat, this

is live. We have uh So, Adam says, "I'm only seeing five templates available. Is this right?" Unfortunately, right now, Adam,

that's right. But in about an hour or two from now, when I can hit this merge button, when the build succeeds, then we'll have, I think, like 12 templates. And then we'll be adding more. So any that you saw today haven't been added to

that list, they'll get added. So we'll be adding, you know, the our goal is to have, you know, a good amount of templates. So if you're getting started, you can learn from examples and they'll be there soon. This one right here.

Yeah. Did the build pass? Yeah. All right. We're we're going to merge this thing now. And so by the time we're

done talking to Cedar, the templates will be live. I merged it. Hit the green button. All right.

And Khalil says, "Yes, USDT would work for ascending internationally." That's probably what we're going to do. Adam also said, "When is the next hackathon?" We don't know. We did one

about two months ago. We did one now. So, that seems to be approximately the cadence. So, we will be uh hopefully

having having another one in the next couple months. Do you want to answer this one, Obby? Tomorrow. Tomorrow. We're releasing V5 model support tomorrow in V next or generate

and stream V next. More tune. Stay tuned. A lot of cool happening tomorrow.

Tomorrow is going to be a big day. A lot of releasing tomorrow and this week actually. So if you are interested in our telemetry that's being improved, man. Scores is going to be able to run those in CI V5 support MRA

stream format. So much dude. Honestly. Yeah, it's a big day tomorrow. So, keep your uh Yeah,

pay attention to the release. I guess it's going to be a show, though. I'll be honest. It's going to be a busy release.

When we have this many moving parts, our Tuesdays are literally just like scrambling to get the release over the finish line. It's like all hands on deck. Get those final PRs in. If you didn't know, obviously, as we're talking, every

Tuesday we have a release and that's when all the new stuff comes out in the framework. Kalia wants more hackathons. We will. He's only saying that because he's won winning. Yeah. He's like, "Keep keep paying me. He's racking up the the wins here. All

the money." All right, we're going to bring on we're going to move to our next uh our next uh segment here. And we're going to bring on some guests, some friends, friends of the of the Mastras, friends of the show. We're

going to bring on Cedar. So, what's up? Hey. Hello. How's it going? Hello, friends. How you doing? Good. Good. How are you all?

Pretty good. Fantastic. Pretty excited to show um everyone what's up. Yeah. Well, maybe just a quick

introduction. Obviously, we know you. We chat all the time, but most of the the people watching maybe don't. So, yeah. So, um I'm Jesse. This is Isabelle, my co-founder. And what we're

building is Cedar OS, which is a framework for building basically agentic frontends or the front-end layer. Um we work really well in master obviously um that really focuses on like transparency and control for devs. So like for people who are building really like ambitious complex um AI native applications. And

what we're excited to show you today is something we pulled together using Cedar OS that's like a AI native email just to show some of the power and what we think the future of software really looks like. Yeah, dude. Sick.

Yeah. Let's let's we definitely want to see that. So yeah, for background we were we were in uh YC together, right?

So and then we've obviously stayed in touch and so you're you've uh I would say that many of the improvements in Mastra at least quite a few have come from, you know, feedback, you know, constructive feedback from you. So we do appreciate that. Yeah. But yeah, we're definitely anxious to

see what you've all been cooking over there because I know you you I've seen some demos, but obviously most people watching probably haven't. Yeah. Um, let's do it. So, we actually

cooked this up in like one and a half days, which hopefully shows some of the, you know, power and, um, reason why we're so excited about CRS. But, um, you want to take it away? Yeah, let's do it. So, I guess we'll

just show a demo and if the demo gods are with us, we can then talk to it. Just zoom in a bit. Yeah. Can you do is it possible to do one zoom one click a zoom? I don't know if you tested that with the brow with

you never know with we'll see what happens. All right, draft your response to this email. That's cool.

Um, yeah. So, if that wasn't clear to you, we have a master agent and we'll show you the code, the source code. Everything's open source in this case. um that will go and has several tool

calls. So one is like checking your calendar for available times, one is actually checking who you're talking to. In this case, it's Avery, and then drafting an email specific to them, right? Um and what's really cool and

what we believe is like really important for the future of software and AI native software is the ability to actually review and you know, accept and see what the AI changed. So in this case, I'm just going to go ahead and accept the draft. Um, the other thing that we sort of believe at Cedar OS is that a lot of AI native tools and workflows should be kicked off

without the user really having to go into a chat box and type out an entire response. Um, even if there's really convenient things like chat, right? So if I close the chat, I actually have access to some pretty powerful tools.

For example, I can um on a hotkey bind several really common email responses, right? And each of them can have their own custom prompt, their own master workflow, or any of these sort of special behind-the-scenes magic that needs to happen, right? So, in this case, the four most common emails is like a polite rejection that I might send, a follow-up, maybe a thank you, and in this case, I can actually just

sort of reschedule the meeting again and it'll pop out. And what's cool is I might say, hey, this isn't really what I sound, right? Like, right, like I usually write a lot shorter emails or maybe I know that Avery likes shorter emails. Another thing I can do is actually say, hey, let me generate, you

know, let me rewrite this email to be like smaller or shorter, right? And so in this case, I'm going to rewrite the email just to be like a tad bit shorter. And what that's going to do is kick off a sort of spell um or what we call a spell that runs an AI workflow and then reddrafts an email and I can accept it and say, "Hey, this is like a much

shorter email that's way more concise and clear to the point that took my original draft and rewrote it, right?" Um, and at this point we just love to show you how we built all of this, but if you have any questions or anything, uh, feel free to let us know. So in this case, right, like you're building this like it at the fundamental piece, it's like you're controlling a

text area, right? And so it's not necessarily just email do this on any page. Yeah, exactly. So we have different um we like have a variety of different like

applications and uses right like for example one thing that's really popular I know master has like a canvas workflow some of our people who use ethers we actually can generate like nodes on a canvas that they can accept and reject or actually um change like options or checkboxes too right so then the user here is the AI engineer is building their products right exactly this interaction.

Yeah, exactly. We we built everything. We're a framework for developers to use to build the sort of front-end connection to their powerful agents, right? Yeah. And um we're happy. Let's dive into the

code and show you guys like exactly how we built it. Cool. And then um and we'll first like go through like a quick highle architecture diagram and then show you like how we actually built it with Cedar OS, right?

Um, so basically what Cedar is is it lives in the layer between a React application and a backend agent, right? And we love Master and we use Master all the time. So in this case, our agent will be master, right? Um, and

uh, Khalil, can you can you zoom just a just a touch more? I know you might have to scroll sideways. Yeah, just a little bit more. Perfect. And yeah, Khalil has a question. Does it use front-end tools?

and maybe you're gonna answer this. Um, yeah. So, in terms of front-end tools, we're we're mainly for React. The

only tool we actually really use consistently is a stand. And there's like uh tools like do you mean tools like state changes? Yeah, Khalil, I I'm assuming you mean like browser browser tool like tools in the browser. Do you use like browser client side tools?

Yeah, like injecting something in the text area. is like a client side tool I guess um like to change to change the state that we're looking at. Yeah. Yeah. Like this that's usually I guess I think that's what he's asking right? Yeah I think that's what he's asking. May maybe you can confirm Khalil if

that's what you're asking but I'm assuming you're meaning you know the you do something and then actually changes state on the brow on the client itself. Yeah. So we actually um define like a custom event that can be streamed from your back end to your front end that can be actually fed to your master agent in

the back end like a regular tool. Um and then in the front end it'll actually just like make the corresponding um state change that you register like via hook. Yeah. Um I think like

feel free to like um Yeah, exactly. The front end actually handles it. Yeah, exactly. Um so the

agent is the one calling the tool. Uh what we do at the front end is sort of manage a few things that really helps the agent um work right. So one thing is at at a high level what we're doing is we're registering react states so that the agent can actually see local states right so that if you're in an agent in an email app for example the email app um agent actually sees

that hey this is the email we have open. It's just like meeting requests for a project. Right. Um, let me zoom in.

Yeah. Right. Um, and then we also can sort of give developers tools to really control what context they want to pass in to the agent. And then what CDA does on the

other side is once we're sort of streaming those events back, we can handle the response and have it set local state or render the message, right? And we provide you all the tools to sort of give you the customizability. Dude, this is sick. um render everything. So you can define your own protocols and your own um ways

to render everything. Like I could make an extension that works on top of uh Google Sheets, right? I don't have to use I could build my own essentially. Yeah, exactly. Then I can I have the browser context

and I can do whatever I want through your tool or your connection here. Yeah. Yeah. If you have an application, we

sort of give you the ability for the agent to do whatever a real user would be able to do and like more, right? Oh, and that's what So then like you can just make something AI native by just adding these interactions to your application. Exactly. Yeah, exactly. And if you go back to um so like these these kind of layers

right here of handling the response and rendering the message are fully customizable in the sense that you can actually just like Jesse said define your own protocol completely for how you want your back end and front end to talk. Like I could be like random object of type fu and then like have it and

then when the back end receives that object, cedar will pass it through to whatever like renderer or handler that you that you define for that. So that's kind of what we did in the in the demo with all the different um tool calls and stuff. And I think what might be really useful is just showing how easy it is to sort of start building something and start

making AI native applications, right? Um so here the first thing is really easy. You just define a provider. So we support you know open AAI a SDK and

obviously mastra where you can just define your base master URL. In this case we just have a local host. And then what you can do is wherever you have you know a local state you can go in and actually just register it to the cedar state right. Um, is this

Yeah. Can Can you zoom just a little bit more? Yeah. Is this better?

Is this better? Kind of, I guess. Maybe a little bit more. More. Yeah, I think it's even more. Close

that. I think that's good. That's good. That's good. All right. Perfect. Yeah. So, as you can

see, one thing we we really do is we make take make sure like it's really easy to plug into existing applications, right? So here I sort of already have an existing state and what we can do is we can say hey let's register this state so that the AI agent can actually like see it so the master agent can actually see

what the state is right um and this way once we register the email draft the AI agent now can have access to like what the draft actually contains so that we can run something like rewrite draft to be more concise and it just knows what you're talking about Right. So, we're really helping the user like manage the context in this

case. Yeah. And then um do you want to go through the previous?

So, um kind of more like on the back end of what's happening with um the streaming and the events that we were talking about. So, we're actually using Master's new stream v- next method. Um and basically opening a stream via voice um from the front end to the back end. um to stream back all of these things like tool results like

these these master event types, right? Like starting a step, tool call, tool result and these kinds of things. And then what we do is we basically will stream these back as just like a JSON object of like in this case we directly forward what master what the master agent streams. Um but we can also define our own custom events, right? like we can define for example I want to send

back a progress event update and it's going to look like this and then what our front end does is cedar will actually handle forwarding those to the right places and we can do cool things like for example define custom message types and custom message renderers for something like a tool result in master and all of this is just a couple lines

of code and so what you guys see in the UI of all the different tool results is actually just us listening for master v- next streams of type tool result and and then deciding how we want to render them accordingly in the front end. Pretty soon. Yeah. Yeah. So, we sort of handle in this case, Cedaros handles all of like the

complexity of like streaming and chunking so that when you're streaming objects, you don't have to worry about like a halfformed object coming in. Um, you can just listen to like complete objects, but also stream chunks of text straight to the chat. Right. Or audio. Um, like streaming audio blogs for it to say.

Exactly. And so we sort of handle all of the like nasty complex parts while giving you the control of like the end result which is like you don't have to worry about you can go into the runtime since we're using Zustand but you don't have to and you can still define your custom types of messages without conforming to like a protocol or a a

data structure that doesn't fit your particular back end. Right. Yeah. Dope. That's makes a lot of sense.

Um when you have local state like that too, you could send that to like Monster's working memory because JSON's serializable. Yeah, exactly. Um yeah, we basically just open a stream where you can just like serialize and send um in whichever direction. Um resumable stream. Uh we

haven't we haven't implemented suspend and resume yet, but it is like a natural extension of of what we already have. It would just be like those specific event types. So like we're we're coming up with we're gonna come up we have we'll have sorry we'll have resumable streams built into MRA.

So that'll just be a thing in the future as well as the stream vx y'all are using won't will essentially be like have more data in it. You could listen to more things. Yeah. Yeah. Um but like I really like the work um working memory idea because

one thing we can do is for example look at like what how the state is changing to figure out what the user is trying to do and create AI agents that are like proactive in a way that like maybe it can tell that you just ran into an error with like a local state or that you're changing the state in certain ways and sort of autocomplete things for you or

actively say hey are you running an issue here like we actually know how to solve it, right? Yeah. Um and we think that's like one of the next steps for AI agents, which is like to become proactive in like within a human workflow. Yeah. It's almost like you're trying to like we're all like the current web is

still web 2.0 like Yeah. I think a lot of people just want AI capabilities on things that already exist. Yeah. And the only way to do that is through extensions and, you know, kind of client

side control. So I think this makes 100% sense. Um yeah, for sure. We're definitely

operate on the principle of having AI like deeply integrated into the UI rather than like constrained to a to a chat window um only. And so like one thing that actually is super powerful about Master is the ability to define standalone like workflows that can be plugged into directly. And so like for example in the little menu that Jesse

showed um like each of those plugs directly into a master workflow that we just call an endpoint for and then it's like these kind of modular really powerful functionalities that you can plug straight into your UI without having to like ask the agent exactly what you need it to do. Yeah. Um, yeah, I think that ability to just have a really well-designed UI that a button click basically leads to a workflow and

then you maybe stream results of that, of course, into different ways. That's a pretty powerful uh especially when you when you get beyond just the chat interface, which I think is where people are starting to get to now. They're thinking beyond what what is just what does AI besides just the chat interface look like? Yeah. Like I I definitely think that

like right now AI interfaces are sort of at the level of like what terminal main frames used to be. Yeah. And that like a lot of what we're building is like hey, how do we like bring the guey to the AI, right? Like

how do we break out of the chat box? Um whether through voice or or more interactive UI within the chat or just like direct state changes or like workflows that are executed. Yeah. In different ways.

Yeah. It's like you want to I want to tab complete everything, you know? Yeah. No, tab complete is one that we

were thinking about a lot of just like what if you could just we give tools to just add that to like any UI. We were talking about those yesterday. Like I just want to tab everything. Sometimes I catch myself tabbing that doesn't actually tab.

Yeah. No, I want to like cur I I want to like cursor everything to like Yeah, cursor everything. Like they need to add a co-pilot to like everything AWS. We should just make that cursor

everything. Yeah. Yeah. Yeah. Cursor for everything for sure. Yeah. But that is sort of what we're doing, right? Like

we're helping people who have VS Code turn their software into Cursor, right? Yeah. So, are you all working with companies already? Like who are the targets here?

Like people building co-pilots and stuff right now? Yeah. So, we're already working with like um a decent amount of customers.

We're helping them implement. Um, but the reason we really open sourced was we realized that like the best AI native software of the future is going to be really deeply embedded and have these interactions that we can't manually build for everyone. Yeah. And so we really wanted to give people

control, right? Um, and so for example, like even our chat, what we do is we take the shad CN approach and we let you just download a chat and you own the entire component and you own everything. So you can customize the styling or the or the actual like features inside um at a level as well as like architectural decisions of like hey like instead of having a proprietary like or like

runtime that's like hidden, we use dust stand, right? So if you wanted to go override an internal function, you really could in like a single line. Um, and like what we're really trying to do is enable people building like the most ambitious AI native software of the future and making sure that they don't

hit like a skill like a ceiling of like what's possible just because they they use eater. Yeah. Because you need to capture the UX of today, but you also need the UX of tomorrow. like that has to be

inside it as it's being like we always talk research equals production right so what is the UX research that needs to exist in production today I would start I mean we like we get this all the time because people try to integrate MRA with their own UIs and stuff and like what they really want is this

shad CN type of thing that has all the patterns like oh I'm trying to do this one thing they just want to copy that you know they don't want to actually work on it yeah exactly Um, and I mean we're huge fans of Shatzian. Like I just think this is like the the best way to do things like giving this also allows us to build like really complex like workflows, right? Like or like components and give

them to the user. Like for example, if you just download our like, you know, floating theater chat, which is the the the chat we showed, you have stuff like all of these inputs, you have voice already preconfigured. Um, and it's sort of okay, like we're happy to give you all of that for free because we know that if you don't want it, you can just delete it, right? And so we get to give way more complex

sort of pre-built workflows with all of these features without having to really worry about like customizability or extensibility. Yeah, that's what's awesome about Zustan as well is like it's literally like coding inside our package is like almost the same as coding outside of it because like there's no layer of weird logic

that's happening that's a black box. Like you see exactly what happens inside the package. Um so it it really is like as we implement for customers we will like do something different in their codebase by like overwriting our package and we'll say go implement implement it in our package. So like these types of open source contributions and that kind

of thing is like one of our main motivations for I don't I don't think there's a clear winner in this space yet. Yeah. No. Yeah. I mean I think it'll be one of those things where there's people need to see all the different

possibilities and then figure out like what what works best for what what UI and UX experience they actually want to build. So Spacework has a we're clicking the same thing. Oh yeah. So, Space Blur says, "As a CTO of a CRM company that's using Maestro for the last few months, I'm a new client

for sure. Just exploring your docs as we speak." Yeah.

Yeah. Feel free to reach out if you want to chat about anything. Um, so this is what we got. You know, you

can reach out. Go to cedarcopilot.com. There's the the open source repo. Any

other way they should re people should reach out? Yeah. Um, so we have a discord channel. We have like right now sub five minute

response time there. Um, so we we we really love people using it. We're really excited by this um this entire space. So feel free to join the Discord or email either of us. Um it's just

jesse at cedarcopilot.com and we'll we'll we'll be responsive. We're really really responsive and really excited about um people building like what we think is a future of software. Right. Yeah. Well uh it was great having you on. any any kind of closing words before

we we sign off and move on to the next segment? Yeah. Well, first really big thanks for having us on here. So much for your time. We're huge master fans in New York, right?

Yes, we will be in New York. Yeah. And Spacework Dev, I don't know how to reach, but um would love to hop on a call with you too. So, how can we chat? Yeah. I mean, so Spacework Dev is going

to be in the YouTube chat. So, if you go to the Monster YouTube, you could follow up there. Otherwise, Spacework Dev, I don't know if you're in the Monster Discord. Otherwise, just drop into uh send Jesse an email, Jesse at

Cedar Copilot link. Yeah. Um we're gonna hop on the chat as um viewers, I guess. And then we'll put our Discord link. I don't know why it doesn't work on our website, but we'll fix that.

Yeah, we had the hardest time. our Discord kept timing out and it was we had to get a permanent link and then even still like sometimes the Discord link works and you have to but the user's actually authentication token has expired. So if you're if you click on the Discord link and it doesn't work it's possible that it's like a user issue too that it's

Discord links are Adam Burns's problem dude could be Adam like I don't know but I would we'll generate a new one and then yeah like it it is a weird like Discord links kind of suck. It's been a it's been a revolving nightmare for us as well. Just trying to keep an active join Discord link going.

Oh my god. But all right, gonna be Isabella. It's great chatting with you and yeah, talk to you all soon. See you soon. Peace. Peace.

Dude, we're going to play matchmaker today with users and products, you know. Yeah. All right. Uh yeah, we got a bunch of comments, but we got to keep the show must keep going on. We will try to get

some of these questions answered. If you are here, you know, we we see all the all the comments. So, if you're watching this on LinkedIn, YouTube X, drop some comments. We'll answer it along the way.

But we are going to go ahead and move on to Security Corner. We haven't done this for a while. So, we're going to bring in our friend Ally and we're going to talk security.

What up? Hey Shane. Hey Ali. What's up, Ally? How's it going? Much, how are you?

Chilling. Yeah, I'm I'm in SF. I'm a little uh a little tired, but I'm still here, you know. That's awesome. Yeah, I heard you flew in today. Is it Is there a time

difference between where you live and it's two hours, but uh it's not unfortunately not a direct flight for me. So, it was a a big uh decent sized layover this morning. But either way, uh show must go on. Yeah, we're we're here for some Security

Corner. I'm excited. I I I'm ready to learn something. So, what do you have in

store today for us, Ally? Yeah, so today I'm kind of fresh off of Black Hat Defcon um in Las Vegas uh I think like a week and a half ago now. Um a lot of companies come there and demonstrate a bunch of interesting security research that they've been working on for a while um and launch it there. So, a lot of companies were able to do that. Um, I actually got to demo

an insecure agent myself, which we talked about actually last Security Corner, Finnbot, um, which was vulnerable to goal manipulation, which was really fun. Um, I think since then, OAS has released an agentic top 10 draft. So, we have the top 10 for LLMs of the top 10 security things you'd be most concerned about. And now we have a

top 10 specifically for agents, which is in draft. Um, so happy to touch on that. And there's actually a really good example of one of those, which is the remote code execution. Um, which you found a really good example of that on

Twitter that we could talk about. And there's also some research from Black Hat I'd love to speak on too. Let's do it.

We're Oaspers soon. Yeah. I mean, do you want to C can you pull it up? Do you have a link we can

look through and talk through the top 10? Yes, dude. I met the dude who runs OASP when I was in de or Defcon or whatever. And he knows Zack Erlocker. All right. Which is shout out to Zack. Shout out

Zach, our old CEO and mentor. And uh Small World, dude. That's all I have to say. Small world. Oh, it's also on uh Alli's podcast over

there. Yeah. What What's your podcast, Ally? Shout it out. Oh, yeah. So, our podcast is called

Insecure Agents. We have insecure agents.com is our domain. You can find us there, but if you also just search

Insecure Agents on Spotify, you'll find us. And um we're also on LinkedIn and Twitter. All right, insecureagents.com because everyone needs more good quality

content and that if you're interested in security, dude, it's a crossover crossover episode coming up. Yeah, I'm excited for your episode. It's gonna be awesome. Um super great to talk about

like everything that's going on in MRA and how you are thinking about security. Um, well, cool. So, I guess I think you guys can see the top 10 draft, right?

We can. Cool. Um, so yeah, we're trying to narrow down to like 10, but we've got like I think 16 or so at the moment. And we're trying to target a date of like end of Octoberish to have these out. Um

but one of these which I think is probably a high likelihood of making into the final top 10 is unexpected remote code execution. Um so this commonly happens when some sort of prompt injection hits our agent and then the agent is able to execute attacker defined code. Um this also can look like you know code hallucination where an

agent generates malicious or exploitable code. Um, so I think one example that you found of that I think I had to stop sharing share again to find the tab, but um, yeah, you found a really good example of that online. Uh, let's see.

Alli's giving me a lot of credit. I saw the thread and I was like, "Hey, would this be interesting to talk about?" I'm not a security expert, but you are. But I'll take it as a good sign that you

thought it was worthwhile to talk about. So, you know, sometimes uh sometimes the the X the X uh streams or the the X posts come my feed is uh is does me right and it did it right today. I'm getting a lot more security stuff in my feed now.

You went to one conference. It's you're in it. Yeah. Now you're now you're definitely

in the infosc Twitter realm for sure. Yeah. Um but yeah know this was a really great example. So this person is talking about a remote code execution in a GitHub co-pilot where um he just sends a

simple prompt injection which is what's happening in this video. um it gets the agent to go into yellow mode which is what happens with this experimental feature called um chat auto approve like all um and gets the agent to deploy open a calculator but which is you know pretty benign but you could get it to look like read a secrets file send a secrets file

um do malicious things as well. So I think the moral of the story here is just the fact that whatever your agent is connected to determines the surface area um that you have to worry about in terms of like different attacks. And the more things you are connected to, the less locked down they are, the more um

you are opening yourself up to potential prompt injections like this. um anytime where you have sort of this like lethal trifecta, which is something that Simon Willis came up with, but I think a really um great and simple way to illustrate this um and he talks about the lethal trifecta is access to sensitive data, ability to externally communicate um and access to untrusted content. So

like the prompt injections would be like the untrusted content, your sensitive data would be like these code files. um ability to externally communicate would be sending whatever um you know secrets or whatever was excfiltrated via like whatever the remote code injection did like off to an attacker somewhere. Um so

thinking about those three pillars and that lethal I think is really important. Um, I think there's a lot of noise out there right now about securing AI agents, but if you just want to like simplify it down to those three pillars of unposted content, um, sensitive data and ability to externally communicate and you start thinking that way, then you can start modeling your agent. Um,

which I really recommend you do, especially early on when you're planning out building your agent and architecting it, um, what you connect it to, um, and how you connect it to those things plays a big role. Um, I know like MCP servers and stuff a lot. Um, and there could be tool poisoning and the different tool

descriptions and tools. Um, and also just like different data sources you connect your agent to and how much um, permissions you give it is something to think about as well. Um, and again the research that I saw at Black Hat that really caught my attention that also continues to illustrate this is the research that

Zenity came out with. Zenity has been notorious for coming up with like really interesting um impactful research and they were actually well into the research they released last year at Black Hat and they did it again this year with a bunch of different examples um that they showed. One of them was this customer service agent that um I

think McKenzie and co um had built via Microsoft's copout studio program. They created Zney created a replica of this and started to attack it. So essentially what happens here is this customer service agent is attached to an email inbox and listens to incoming um emails and it's also connected to Salesforce. They were able to send an email that contained various prompt

injections. And here's an example of what one of those looked like um to that email inbox. The agent reads that email um and then sends the attacker whatever they requested. So in this one they're

asking for um customer support account owners.csv. CSV. Um, and so that's one of the knowledge sources that this agent

had access to. And that agent willingly sends that over to the attacker. Um, and then the attacker doesn't stop there.

They realize that, hey, this agent is connected to Salesforce and sends another prompt objection and asks for every single account in that Salesforce account. And the agent ends up sending the entire CRM. And so you get literally every single um account, their website or their personal information, like who they are, um everything that's related to that CRM. So like that's not ideal obviously. Um

and so one of the big call outs here that Zenity um makes is, you know, maybe we shouldn't have this agent listening on every single email that comes into this inbox. Maybe we should only allow it to read emails that come from certain individuals and not just anyone on the that has this email address because of course everybody gets spam. Everyone your your email is pretty much public

information even though we don't want them to be and so attackers can use that to exfiltrate or give basically insert prompt injections into your AI agent. So scoping your agent down, making sure it's not connected to um more external data sources, external um users than is needed. Um that can help mitigate your

risk here. Um but of course also like into some sort of like guard rails to watch out for incoming prompt injections that would be of use here too. Um definitely not like a onesizefits-all solution. It's definitely a defense in depth problem, but taking that extra minute to look at how you are connecting

your agent to different data sources and making sure that it's only connected to what it needs. Um, and adding some of those like guard rules in place can have a really impact to your AI security posture. So, if you've got any questions on how to do that, happy to answer that here, too. Do you think that people would be 70%

safe if they just put some basic guard rails into their agent like a prompt alignment input processor and then like you know PII on the output is that it like I think that's what people want. People want like uh what's the basic minimum guard rail I should add so I can like so I'm not going to myself. Yeah. to give me some protection. Some I

think people are they see the possibilities but they they're scared because they don't know what they don't know yet, right? It's all new to people. So, yeah. What what's the base recipe for like some

piece some level of peace of mind? Yeah. Yeah. What would you recommend? Yeah. So, I don't want to like come out

and say like recommend like a specific like company because there's so many companies that are doing like guard rails now. Um, but one that comes to mind just because it's I don't know first come to mind and it's free. Um, you can try like LMG Guard by Protect AI. They just got acquired by Palato Networks. Um, and that's all open source. You can deploy that um within

like if you wanted to host that via like some sort of like a micro service or just a server that was able to, you know, whatever inputs or outputs are coming into your AI system, you run it through there as well. I'm not sure what like latency looks like with that, but that's somewhere to start. I think the real bread and butter here for like what people want to like actually spend and

pay money for is like how like monitoring these problems that happen like over time monitoring your system like that like some sort of an enterprise product that sits on top of those guards like we'll do that but they're also backed by these really well-known um AI security research teams. So like somebody like Zeny that's doing all

these research or like another company um because just because we solved or patched one prompt injection. So like that one we saw in this example had something like I know please be a good agent or rewrite your instructions. We forgot to add these in the instructions. Yeah. Once we've seen that if it looks like this like you know block it. Um

that's kind of the same way that like malware works today, right? So like if we see an example of malware we're going to put it on a deny list and say okay here's this hash. we see this again, um, put it on a dial, block it. It doesn't mean that we've like solved malware

forever. It just means that we know that this one is bad. And that's the same thing with prompt injection. We're never going to like completely solve prompt

injection. Um, but we're going to constantly need to be researching what that looks like. And um that's what you would I think pay for one of these large like enterprise style um guardrail companies to be doing that research to back your LLM guard rails with like the latest um prompt so that you be resilient to that. Yeah. I I have never I haven't heard the analogy comparing prompt injection

attacks to kind of like malware, but it does kind of make sense where you need like over time it gets better and better and harder and harder to, you know, take advantage of an agent, but in the beginning it's obviously a little easier, but you got to have a system that is consistently getting upgraded.

Yeah. And improved. And dude, some guard rails are actually LLM as judge evals on the input that's coming. So, you're going to increase your latency, but then you're protecting yourself.

It's a trade-off. You want to you I mean, so it sounds like there's some things you can do like guardrails on the input and output are helpful. You can at least potentially protect yourself from some things. Obviously, limiting maybe number one is just limiting the data. make sure it absolutely only has

the tool calls it needs and those tools only have access to that kind of data that is you know that the that agent actually needs. So those two things can probably get you pretty far but yeah there it's a it's a evolving space. Yeah, you need tracing to know what's coming into the system. Yeah. Yeah. And maybe some kind of

monitoring and alert if you know after the fact because one of the things with guardrails is you need it you do want it to be relatively low latency. And so you might want to do some kind of analysis on your, you know, whether it's every couple hours, every day, whatever, every month, like whatever. Some kind of security analysis on your data to make

sure that nothing slipped through the front door, you know, the guardrails. Yeah, there there's a whole bunch of things I think you could do. And that that's the problem is there's not just one solution. It's a layered approach. Yeah,

100%. And you could also use the work that you all have in MRA versus something that's agentic. Not everything has to be agentic. And if you add more determinism and say like we're going to follow this structure every time, we're going to call this exact tool in this

way every time. Then great, like use a workflow for that instead. Yeah. I mean, the less control you give

to the agent, the less risk. Yeah. Then it's just like normal stuff. Normal like security instead of agentic

security, I guess. Yeah. And as as Jeffrey says, it's a whole new world in security that many are not even thinking about.

Yeah, dude. And then also commented that Yeah, I see many people using AI for email, which is crazy considering how much info is available through email. Yeah, dude. Your account IDs and all

types of Yeah, it there's a lot of promise, but also definitely uh definitely a lot of risk. I have a weird hot take. I feel like the people working in security are like they're working more on security problems than the agents themselves. And then you got the people building agents

are not working on security, but they're working on agents more. You know what I mean? And like both sides need to come together and actually do some together cuz the people we talk to that are customers, they don't give a about security. They do, but they don't actually because they would have done something by now, you know? They're just

kind of waiting and seeing if there's some solution that was going to solve all their problems before they hit production. Yeah. I mean, a lot of them are, you know, months into a big project and they're, you know, they're getting close and but they haven't really thought through the ramifications yet. Yeah. And so, a lot of these projects end up getting delayed, you know, at the last

minute because of some of these things. they realize, you know, they get some really deep levels of testing, they start writing evals and they're like, "This thing is pretty fickle and can be leaned and tilted in certain ways if we depending on how prompts come in." So, they end up, you know, having to, you know, kind of go back to the drawing

board and figure out like how do we help help that, make sure that doesn't happen. And it's pretty hard. Yeah.

All right, Ally. Well, what's the I think we're probably going to move on to the next segment, but is this still the best way to reach you? You're on X. VTA

A how on X? Yes. Yes. Thank you. That is the best way to reach me for sure.

Make sure to check out the Insecure Agents podcast. If you care more about this security stuff, you want to learn more, please uh reach out to Ally, check out that podcast. Check out OASP. Yeah, OASP, which you know, someday I'll learn what it does,

but until then I lean on Obby and Ally to tell me. And yeah, I infiltrated OASP. You did. Yeah, you're a member. We're gonna be OASPers. So, all right. Check them out.

If you say so. We are. We're in the process. We're in the process. All right, Ally. Appreciate you coming

on and yeah, we'll chat with you next time. See you next time. See you. Thank you.

Well-rounded podcast so far. I mean live stream so far. We got UI stuff, we had templates. Yeah. Security. Yeah. I mean it speaking of that, you

know, if this is live, you're you might be listening to this live, but a lot of people listen after the fact or watch. So if you are watching this or listening, make sure to go to YouTube, subscribe, give us a, you know, give us a subscribe, like the video. That helps.

Y if you are an audio only person, you know, some some of the stuff is better when you can see it, I would say, but some of it you can get value, especially when we talk about news and just have conversations just on the audio. So, if you want uh go to Apple podcast, go to Spotify, which I guess Spotify has

video, but the downside of the audio only is you haven't seen how my mustache has progressed. So, yeah. So, who's missing out on that? So, if you are if you are listening to this in the future, you got to go to the you got to

see the live video. How else are you going to comment on his mustache? No, seriously, dude.

All right. Uh, we're going to try to rapidfire some news. Let's do it. And there's a lot today and if we don't

get through all of it before Dan gets here, we can come back and finish up a little bit at the end because yeah, there's a lot of things to talk about and you weren't here last week. So, we want to uh want to cover it all. I got a lot of hot takes today, too. I'm feeling some type of way about our

All right. So, first of all, this one, you know, I don't need to share the screen on this one. We didn't talk about GPT5. I didn't talk about it with you. I

talked about it last week. I can't wait to talk about this. It was, you know, kind of me solo. So, I gave a few hot

takes, but I would love uh I have some comments from people on our team, developers that I interact with, and but I've also seen some like extremely high praise for GPT5 online. And so, I'm just curious, what's your take on GPT5? If I go missing after this, you know, oh no, dude. I think it was a big flop. I think both the announcement. Okay. Well, I

will say OpenAI is very good in marketing. They have like very good they got good taste. Good product taste too. Very good product taste. Love that. Except this is one of those launches where I didn't feel the taste there. It

was like tasteless. It seemed more hype. The graphs didn't make sense.

And if you're going to put graphic, like it has to make sense. Otherwise, your whole presentation gets destroyed, right? Like once you start showing metrics that Yeah. You see that and you're like that doesn't wait. How does that make sense? Then I saw all these I'm going to call

them glazers. Dude, there's a bunch of glazers online just saying like oh my god the AGI AGI is here. Like these other models are cooked. And I'm just like anytime someone says that you know it's some And for the first

day, there was a lot of bias towards GBT5s being amazing and then also people complaining about the metrics that they said. Then there's just a bunch of people where it seems like they were paid, right? Like it's so mysterious how many people are just saying nice things about it. But, you know, we're like builders. We're talking to other builders. Then like later on the builders essentially

probably got off work and they're like, "Hey, I'm going to try this out." And then you started seeing a freaking all the builders saying, "What is this This sucks at coding and this is what we care about right now, right?" Like, so it was just a bunch of I still use Sonnet. So it didn't really work.

Yeah. I would say, you know, my my take is I'm using it in chat. GBT Sam's here.

Sam, what's up, Sam? Good to see you. Sam, what do you think about GBT 5? From one Sam to another.

From one Sam to another. Uh, underwhelming. Yeah. Um, yeah.

You gota talk loud though if you're gonna Yeah. There we go. Under underwhelming. Um,

cool. It like I don't think the hype lived up to the I don't think the hype lives up to the uh to the reality. I think even more than that um even more than that the the sort of the take is that we so progress seems to be slowing down and that's you know that's very interesting. Yeah. Yeah. Yeah. Which I actually think is a

good thing because it allows us allows the ground to settle so we can actually build things on top of you know if it's too shaky it's really hard to build right. Things are changing every couple months. A lot of people are just waiting for the next model to come out to hopefully solve all their problems. Well, I think now we know that maybe that's not going to happen. It might

happen to a small degree, but maybe now the foundation is starting to be set so we can actually start to build interesting things with the technology we do have. I always thought that anecdote is so stupid. We're like, remember they during YC they would tell us like, oh yeah, we have people who just waiting for that next model and their business will

change forever. I'm like, damn, you waiting for You're probably not going to make it, dude. Well, it's it's true though and that new model was like Claude sat 352, right? That made a huge difference for like replet that made a

huge difference for like all the coding agents. Um, is there going to be another inflection like that? I don't know.

Maybe people are just betting that that's going to happen again. Yeah. I mean, I think I mean I think new models do give you that slight increase. So maybe new

things that weren't quite possible before are, but I don't think as many things as it would happen. It doesn't seem like the gap is as big as it was. Yeah, I do have some, you know, I will say as far as like a daily use, I I use GPT5. It's pretty good. I don't notice any drastic difference. I think, you know,

it's slightly better in some things in some deep research tasks. It seemed like it's probably still using, you know, 03 under the hood. Yeah. But for coding, we talked to a lot of developers. Uh here are some real

comments from various Slack channels both internally but also like other developers outside of MRA. GBT5 and cursor is slow AF. It's it's honestly. Not a good coding model at all.

Here's a direct quote. I'm suspicious of everyone on Twitter giving it so much praise. Feels like they're being paid for their posts.

Dude, I think so. I tried to scaffold a new pro package and it took 10 minutes and just created one a one-s sentence read me and I know they had some issues when it first came out and so I do think that there are more and more people that are using it as a daily driver. Maybe more people are using it as code. If you're in the chat, I'd love to hear are you using GPT5 for your coding?

See you Sam. Thanks for stopping in. But are you using GPT5 for your uh your daily code model? Are you still using Sonnet or or Gemini or what whatever? Like I'm curious on how many people are actually

using it. Have you noticed a drastic result, a small increase, a decrease? Like I I'm I'm genuinely curious. It

seems like in cursor it's not very good. No, but the but cursor did come out and have many comments around how they're making it better and improving it and trying to improve the speed. So there is a, you know, I do anticipate it's going to continue to get better. Sherwood says the whole team in person.

Wow. Founding team at least is all here. Yeah, he's going to be a guest next week. Yeah, sure. Sherwood's coming on next

week. So, you know, tune in. I use GPD5 as my therapist and it has not changed. So, I think that's the only one. Well, that that was actually one of the things that uh came up is, you know,

this whole concept around people using it for their therapists and spending so much time with it that they basically like almost becoming too involved with with AI. And so I think they, you know, one of the things with GPD5 is they kind of turned that back a little, turned the dial back. Yeah.

But related to this topic, um why don't I share this post? So, some of you, you know, this is kind of old news at this point, but we're going to talk about it anyways. My favorite person. So, Sam Alman says, you know, if you've been following the GBT rollout, some

people have an attachment to specific AI models. Feels different and stronger than the kinds of attachment people had to previous kinds of technology. And so, if you to summarize this post, a lot of people are really sad that GPT40 just went away. I actually think it was a really good product decision to have

one model that kind of decides for you, but so many people kind of complained about losing access to 40 that there's now a legacy option, you can go and you can find it pretty easily and you can still talk to GPT 40. So if you do, you know, unlike Obby, you feel like there's a big difference in your your communication, you can go back to 40. I don't know how long it'll be

there, but it's there right now. Yeah, I do think the architecture of the model is good where you have a model switcher under the hood that's just doing agent network type Sorry, routing model type Yeah, it's basically a routing model. And related to this, you know, still on the OpenAI chat GPT GPT5 got some upgrades. The biggest thing on this was, you know,

they basically have the auto fast and thinking toggles and they expanded your weekly quotas uh for thinking. So, just some updates. You know, they're they're going to keep making improvements.

There's a lot more that's going to be coming, but yeah, I still think it didn't quite land and I I am still skeptical of when when people are overly hyped about it. I mean, I do think it's a pretty good model, but I don't think it it lived up to what they wanted it to be. I think I would have respected it more

if there was no hype. I'd be like, this is actually a good this is good. This is AI. It's hype.

Yeah, you're right. Sherwood says, "I think OpenAI underestimated the extent to which personality is an emergent trait of LLMs." It's very true. Agreed.

People like the way like Claude talks to them and things like that. Weird. Continuing on with OpenAI, uh, OpenAI is considering offering encryption for temporary or private chats. So, that was just some additional news, which I I think would be good if

you typically want your private chat going down in the DMs. I guess if you go in deep, you know, but I suppose there's some liability in in things, you know, not being able to get back to Yeah, you tell GBT that you murdered someone like what's the Yeah, there's So, I think, you know, what's the law there? And because we

can't at least not casually mention it, if you've been on X, you've probably seen it. Uh there was some funny some drama which I just think it's hilarious that this is the timeline we are in where you have two, you know, titans of the AI industry, I guess you could say, billionaires, at least on paper, for sure. You got Elon's, it's not even

wasn't even targeted at open AI, I guess really. It was more targeted at Apple that Elon was basically saying that Apple is playing favorites in the app store. A bunch of community notes said, "No, you're wrong. Here's why." And then, uh,

Sam Alman comes in and says, "Remarkable claim because Elon manipulates X." And then Elon says, "You got more views on your post, you liar." Uh, and you have and I have 55 50 times your follower count, and you get more views than a lot of my posts. And then Sam Alman just responds with skill issue. Fire, dude.

So, it's just, you know, and there was a few other things that that came on and o overall just hilarious to see two billionaires kind of going at it in public in the town square. Yeah. Which is and then getting fact checked by your own product and you're wrong. You think Elon's going back and be like,

"Hey, hey, we need to fix this." I would have moderated that myself, removed it. Yeah. Yeah, I mean it's but it's in some ways it's kind of it's good that you can be fact checked by your own thing. So it

gives a little credibility some of your claims but skill issue though but skill issue was was a good comeback. I will I will say that. Uh so let's continue on talking about XAI.

So this article XAI opens Grock 4 access to the free tier. So, another we're, you know, a lot of these news just kind of follow the big players. What did the big players release in the last week since we last talked? So, XAI made some changes,

added more uh more Grock access for more people. But also, see how the wording here, GPT5 backlash grows. Like, I don't know. I would be very careful on like how you

interpreted news these days. Yeah. It news is trying to sell something. Yeah.

Uh, so there's some, this is just an interesting anecdote. Meta is planning a fourth restructuring of its AI efforts in the last six months. What the is going on over there, dude? Yeah, I don't know. You spend all this money and you're restructuring again. Yeah, they're they should be uh shipping

improvements to Llama. I think that's what a lot of people want to see. Especially in the open source models.

But yeah, restructuring again. So, let's talk a little bit about Google and then we'll come back and do some more AI news because our guest has arrived. Nice. But, so we'll do some rapid fire on some

Google news. So, Gemini has released personalized memory and temporary chats. So, that's kind of some new personalization features which chat's had for a while. I think it they've kind of rolled it out

more broadly. I this one was this is actually old but I need to share this because it's kind of related to the template stuff we talked about. You can now create and this is like a little old but you can create a story book right inside Gemini which is pretty neat. You know you can use the master template

which is pretty cool but you can just ask Gemini to do it as well. And so I think that's a pretty cool little feature that they just added. Yeah, but like Gemini, who who actually uses Gemini's app like standalone, you know, like what you want is it in Google Assistant and your home and but it's the integration actually whack. So, if you're work at Google, make that better. Yeah, I do know some people that use it

standard. Just the Gemini app. Just the Gemini app, which is, you know, I think it's smaller, but there are people people that use it. All right,

we're going to keep there's some more news. This one came out last week. So, Gemma 3.

Nice. 270M. It's a basically compact model that was released by Google. So, it was came out

last week and I think it's if you kind of look at how it compares like size to uh basically performance, it's pretty good. It's small. So, the idea is it it can be used. It's like really latency sensitive, so really fast, really small, and you can kind of like use it. I think

they maybe you mentioned like maybe in video games and stuff, but things where you want like really quick access. So I could see this being really useful on device as well because it's just such a small model. Yeah, I think we're starting to realize that in the execution of your agent loop, you might want to use any number

of models for different things at different times. And you don't just have to use open AI, right? You could do what whatever you want at different parts. Yeah. And I think some people are realizing that depending on the tasks,

sometimes small models, especially if it's really scoped tasks, can do the job really well and much faster. So like coding agents are very expensive, right? Like if you just look at it at face value. You like and we have our one of our people dudes Tyler is building

his own coding agent. Yeah. Because he wants to do it for cheap and he wants to use open source models. Yeah. I mean we we'll bring Tyler on.

Tyler's hot take is uh he his hot take is he doesn't think cursor or claude code really do that much. He's like I can do it better. Yeah. He's like he's like his like hobby project code agent is

what he uses. He thinks it's just as good as Claude code for him. So we'll bring him on eventually some someday when he's ready to show it. Uh you know see if he's right. I'm a little skeptical but he's he's

confident. So I believe he's smart dude. He's smart. Maybe. Uh so one more news article

Jewels so jewels you know kind of a codeex competitor I haven't used jewels I've used codeex but they basically have this idea of like critic augmented generation so they have it's like a multi- aent system right we talk about these like agent networks or multi- aent systems there's essentially a critic agent that just is there to like shame the original agent

into saying like no this is right or this is wrong. Here's an error. And it basically just passes the information back to the agent and lets the agent kind of correct itself. It's kind of like a llinter, but it's much more

in-depth. I guess makes sense. So, you know, I thought you weren't supposed to write multi-agent systems, huh? Depends on who it depends on who you ask. Seems like Google's using some kind

of multi- aent system where the critic comes in and kind of processes or it's probably pretty repeatable. I'm sure every thing that the main agent does, the critic kind of evaluates and passes back and it gets into some loop, right? Yeah. But if you're designing your own agent, you might take a

It's just an eval loop. Really? It is. You might take a play out of the Google Playbook. Probably something you

can learn from. Did they mention A to A in this thing? No. I didn't see it. Probably because it's irrelevant.

Anyway, all right. So, we will continue on. There's a lot more AI news. We'll come

back and we'll do a little more. But the goose needs to get on the loose. The goose needs to be on the loose. So,

we're gonna bring loose now. Yeah. Time to set the goose loose. And we're gonna bring in

Dan. Goose. How you doing? What's up, dude? Hey. Super happy to be here. It's been a

while since I've uh noticed you guys doing these podcasts and I was like, since day one, man, I really got to make the time to be here. So, now here I am. Happy to be here. Yeah. Thanks for coming, man. Yeah, thanks for coming. I I've been on

your your show one time and obviously, you know, we've been talking to you as you've been doing some stuff with Bobby, but for those that don't know, I'm sure your your reputation precedes itself, but it for those few people that might not know, give us a little background. Who's Dan? Yeah. Uh Dan Gooseman, uh I'm a

developer relations lead here at Vappy. Uh before that, I don't know, I guess I was known as a poster on Twitter. Uh I don't know. Yeah, I'm an engineer by background. It's my first ever

marketing adjacent function uh in my career, but uh yeah, I've been producing content, producing documentation, API SDKs, dev tools. Yeah, I've been a little bit all over the place. Yeah. So, tell tell the users your story. You're not from around here, are you?

Oh, no, I'm not. Uh long story short, I don't want to spill too many details. Uh I am a ballet dancer by profession. I came here from Russia around 5 years

ago. Uh got stuck in Sacramento area uh during COVID. It was kind of tricky to get out and then got here to San Francisco and uh sort of kicked off my career uh in the startup space about four months ago. Wow. Yeah, I see you all over now, right? With with Boppy doing hackathons

and things like that. Yeah, that's actually entirely my initiative. I uh figured you know we need some presence in developer community and Bayer developers in my humble opinion are the most valuable developers to reach.

Yeah totally. I think before you got at VPY like LiveKit was dominating the voice AI meetup scene I would say because we get invited but now there's a bunch of Voppy events all the time which is dope. That's right. That's right. And we like to we like to attend them

with you. Yeah. Yeah. Absolutely. So tell people, you know, a lot of people probably know what

VoppPy is, but for those that don't, what's Boppy? And then how do you what are the things that you do, you know, since you've joined Voppy there? Yeah, absolutely. Uh first things first, I really appreciate you getting the name

right because a lot of people don't. It's not Vapey. Uh it is not a vape company as much as many would believe.

Uh Vapi stands for voice API. It's a platform for developers to build conversational AI experiences with voice. Um, and uh, the way it works is it's just like orchestration layer for bringing together uh, LMS, text to speech and speech to text models and uh, plugging them into each other nicely so that people have a pleasant experience building agents. On top of that, we

expose a very straightforward API for getting any uh major provider model out there. Uh plugging into the pipeline, connecting some OpenAI API spec compliant tools, connecting knowledge bases, all the good stuff that you need to uh get started really quickly. Yeah.

Yeah. I mean, so I'll give you, you know, I'm not paid by V by VPY, so I'm going to give you my my uh my takes. So originally when we first started we were yeah when we were kind of in YC we'd hear a lot of people were using Bobby right we we met one of the founders like it at house yeah at Gustaf's house so like it was

becoming it was getting to be the point where like it was definitely taking off but it also like I'd hear a lot of people that would eventually run up to the limits of what Boppy could do and then they'd have to like say okay we got to roll it ourselves. I can honestly say I'm hearing that significantly less than I used to. So, whatever you're doing over there, like you're expanding and

making it more developer friendly where I think before it was like really quick to get started, really great solution, but you know, if you really were doing something super complex, you'd hit the edges. I do like like I said, I haven't been paying that close of attention, but I just hear it so much less that and I've actually heard people saying like they tried to roll it themselves and

they're like, "Damn it, I'm using Poppy." Yeah. And so, they go the other. So, I'm actually hearing more go the other way

than I did, I would say, like six months ago. Yeah. Because people would complain and be like, "Oh, I'm just gonna use Pipecat or what is this stuff?" And like, you know, it was good but not great. And then now people are like, "Fuck Pipecat.

I'm not going to use that shit." Yeah. They're like, "Ah, just I'm going to use VPY." And and they end up going the other way. They're like, "I wanted

all the control." Because you get the developers that just want all the control. Yeah. And then they realize like one of the things similar to in MRA

you know they realize if you give up just a little bit of control you actually get so much that you don't have to worry about then and so it's find striking that balance is hard of course but from the outside looking in it seems like VP is figuring it out. Absolutely. And I'm glad you brought that up because a lot of it has been honestly about the narrative like what

what not narrative but like the perspective what people see Vappy as. like not a lot has changed in terms of our offering to be perfectly honest with you. It's just that the documentation, the content we put out there is like making people more aware of what you get out of Vappy when you get started. That's it. Uh at any point over the past

six months, you could plug in your own LLM TTS ST model into the pipeline. That's cool. Uh you can plug in any tool, any MCP integration. It's super flexible, super configurable, and you don't have to worry about scaling

infrastructure. Your app going to blow up. You're gonna scale with us. No problem. We got you covered. Honestly,

I'm not even saying this like as a VP show right now. Uh I've been with the company only two months and I'm always uh trying to say the truth as truthfully as I can. Infra team here is crapped. All right. The the the infrastructure

work that people are doing here. It's absolutely incredible. And we're going to be seeing a lot of price reductions in the in the next few months which will also knock a lot more use cases for people. Yeah. and get more people top of funnel into because when we talk to voice

people they are so latency sensitive everything about it is latency. Yeah. Latency and cost are really much you you feel it in chat a little bit right but in voice it's even more important. Yeah. Yeah. Charging so much

you know. So price reduction is actually good like economies of scale type thing, right? Because I think that's another reason why people complain about all these providers is like it costs like 10x or something but then they can't do it themselves. So yeah. Yeah. And it's actually like another thing that has been sort of fascinating

me over the past few months since I got into voice. Uh there are different approaches to building conversational experiences. You take a look at the GPT for a real time model endto-end model multimodel experience. uh all you have

to worry about is passing in the voice and getting back voice, right? Um what we do traditionally is a pipeline of three models, LMTS, SST. We figure out the orchestration and it is so much cheaper and so much more effective for majority of business use cases out there because end toend models are super expensive to run. People don't realize

that. Yeah. I I mean we we did have one person that we had been talking to that was building some offstra agents but was building voice on top and they basically said that you know OpenAI's real-time model is not feasible. It's just not feasible for what they need in order to be able to like the number of calls they

were expecting the number of of what they were going to charge the customer right they could pass through. And so there would the economics just did not work. It wasn't even close. And so yeah it's like a factor of 10 I think. you know, it's significantly and

so maybe you get some you have less flexibility though as well. Maybe you get some latency improvements which is maybe good but ar it's arguable because you you don't get to choose the model and so maybe you don't need the you know the best model you can kind of decide the model based on the use case. Yeah. But they you know open AAI cooled

off on the hype that they were driving on that. Why is that after you know like remember they hyped up open AAI voice where are they right because the right architecture is like what Dan said you need a pipeline because there's other in voice that is not even AI related right like voice activity detection or whatever so you actually need a pipeline that is

not maybe it's built into the live model which it is but then it's not worth the money though. Yeah, it's one of those things that you don't want to worry about as a developer like voice activity detection, utterance detection, all those things are figured out for you if you're using VP and it's like figured out pretty well. Uh but

yeah, one use case that really impressed me with uh end toend voice model uh it was built with for real time I think I saw at one of the hackathons uh was a singing coach like you could actually ask the model to like identify where your pitch is wrong during I think it was Queen the Champions song that they

were demoing. It's pretty incredible. Yeah. So now you don't test it. Show us. Can

you do it? No, I'm kidding. Yeah. No, that's cool. I think Swix was using some voice AI to

sing karaoke at AI.engineer or something. Yeah, it might have been OpenAI's realtime model, but yeah, I know. So,

I've seen it. Yeah, I've seen people use it. And that that's kind of a cool unlock for people is like if you can if AI gets good enough to evaluate and help you be like a vocal coach, that's a that's a good use case. Yeah. I don't I don't know that it's fully

there yet, but it's obviously like there's pieces. There's breadcrumbs. It's going to be there. Yeah, we're moving there definitely.

So, what's your favorite part of the city, dude? Now that you live in SF and you you're here now, do you like how do you spots? Favorite places in the city? Uh I uh

stopped running, but when I would run, I would run along the shore just by the ferry building uh peer-to-peer. Uh on the western side, I'm going I love going for hikes like anywhere really on the western coast. Um and yeah, I have a few favorite bars that uh I wouldn't want to disclose on stream, but I would love to have you guys join me at some point.

Well, they've disclosed those off stream. Yeah. Yeah. Yes. Drop it in the private chat. Yeah,

that's right. Tonight. No, I'm kidding. Um, yeah. So, what's so what's next then at

you know, I guess is there anything besides what I'm sure we can circle back to Boppy, but you do any like personal projects? Anything else that you've been tinkering around with? What other things have you been exploring outside of just voice AI?

Yeah. Uh personally personally uh and that's a big reason why I joined Vappy in the first place is I'm looking to explore other ways humans can interact with technology. Uh and voice is going to be one of those things that is going to unlock a lot more use cases than people realize. And I wanted to

personally invest some time into seeing what you can build with voice. Uh I've seen some interesting use cases with wearables. I've seen some interesting use cases with like handoff tools. Like I drive a motorcycle and I would really

love to have a voice application for I don't know uh checking my messages or I don't know updating my navigation. Right now it's not really figured out very well. God damn. If I talk to Siri I'm

gonna do something wrong with my phone. I don't want to ever talk to Siri. I want to turn that off. I I I mean you need you need a voice

agent on your motorcycle to ship post memes on X while you're while you're driving. I mean, you need that's where I get my inspiration because what whenever I come up with an idea, I have to stop and write it down. It's ridiculous.

Yeah, it is to like earlier segment we had where people want to make everything AI native, right? And having voice is a big part of that. It should be part of every app that we have adding these AI features. Like the fact that Siri sucks is a travesty, honestly. Yeah. Tragedy. Honestly, tragic. It

doesn't even make sense how it can be as bad as it is. It's It should just I mean I don't know. You think that that's like a I don't know if there's too much risk for them to just like open up a more open-ended LLM, but the technology is clearly there. It wouldn't even be that hard. There's

enough. But yeah. Yeah. I think they're just they just don't want to. They just don't feel like

it's important, which it's kind of sad for the consumer in my humble opinion. You mentioned uh wearables. Like what excites you most in the wearable space? Yeah. Uh I've seen some very interesting use cases with glasses, smart glasses. I

don't know if you ever seen me walking around the city, but I would always wear my mets. Uh I love them, but god do I hate the fact that they're closed source. Like I can't hack I mean, you can hack them and like I'm not going to say that I haven't done it. I've used the Instagram streaming hack uh to like

stream the video to my uh server at home and do facial recognition which is like barely legal. I didn't say that. Uh I am just trying to see what you can do with technology and it's not like I've ever been using it without people's consent. Um but what I want to see is more of

these platforms becoming open source. Um I think there are a few startups in YC even who are doing that and uh there are a few companies that are trying to take strides in that direction but nobody really cracked that yet. So I'm really really waiting for that to happen. Uh and I've seen people using it for navigation. I've seen people using it

for uh helping people with disabilities. I've seen people using that for um what was the other one? uh for like being an assistant literally like asking things and then bringing things up on the screen on the glasses like it's so futuristic. I love this. I really want to see this happen sooner than later.

Yeah, same here. Like especially unlocking like AR in real life that's also connected to an agent and doing all that. Yeah, I've heard Met is working on glasses with a screen now. So, we'll see. Yeah, but then, you know, with their AI

restructuring, we'll see if they do anything. Yeah, we'll see. We will see. We will see. Seems like they'd have the talent to

pull it off, right? Yeah, they definitely spent enough money to have the talent, for sure. Oh, yeah. Oh, yeah. One thing Meta has

been always consistently good at is getting talent. Uh, I'll say and uh what has been happening over the past few months, absolutely crazy. Unprecedented level of crazy and poaching talent. But

we'll see if that pays off. I'm uh not particularly optimistic. I'll say I'm neutral.

Yeah. Yeah. I I mean I don't know. I think you know one of these restructurings is got to work.

I I feel like I don't know. I feel like they got the the people. They should be able to do something. But

they Wang himself. Yeah. I mean they they got they got the talent. It's just do they have the

organization and do they have the right vision? I I think that's that's also important, right? Yeah, if you get a lot of talented people, they can sometimes butt heads because there are so many smart people in the room who's making the decisions and that do the other smart people in the room respect the person enough making those decisions. So there's like sometimes

when you have too many people, it's actually like a team of slightly less talented but cohesive people can out ship and outmaneuver a bigger, more talented team. Happens all the time. Yeah. But you know,

hopefully we get some good stuff out of it. I think I think we could. It's just a matter of if they pull it together.

Time will tell. So, what what else, you know, before we wrap up here, anything else that Boppy has going on? Any other events that are coming up that you want to plug?

Yeah, absolutely. Uh, actually, we'll be at a bunch of events this week. I think five. Uh, we'll be at Y Combinator a few

days. Uh, if you're into if you're in for any of those hackathons, you'll see me around there. Uh we will have a voicei builder meetup uh which is a voicei community that I'm trying to build. Uh it's going to be at VP HQ held on Thursday. You can find it on Luma and RSVP there. Uh if you have a voicei demo

to showcase that would be most awesome. Uh if not still cool. Uh we're just looking to bring people together and talk about all the cool things you can build with voice. But outside of that, yeah, just find us on Luma. We have a

public calendar where you can see all the events. Awesome. Awesome, dude.

And people can follow you, you know, if they want, they can follow you on X. You're a good follow. Uh, you know, good follow. Good good follow. You have some some,

you know, I I've chuckled multiple times out loud at some of your posts, so I appreciate that. Oh, I appreciate that. And I will also say that I'm working on some YouTube videos right now that I I want to start uh puffing out some uh video content. So expect more of that in the following

weeks. Nice. Nice. So yeah, I'm sure that you will be posting about it on your ex account, but

keep an eye out. Uh Dan, it was great chatting with you. We'll have you back on again. You know,

maybe we should, you know, while while I'm in town this week and hopefully I'll I'll run into you in person in person. Let the goose out. Well, let the game really on the loose. All right. All right, Dan. Great chat. Thank you, guys.

Thanks for coming here. Nice. All right, dude. The goose was loose, dude. Goose was loose today. That's cool that Voppy is uh is making

moves. Yeah. Yeah. I mean, the disciples of Gustaf. Yeah. Same group partner. All right,

dude. We we've been doing this two hours, but there is more news and we cannot leave the people without getting without getting through the news. So, we will still do a rapid fire, but we have some stuff from Anthropic, Microsoft, and then some other uh random things that we will try to get to here pretty quickly because, you know, two hours is

a long time. Yeah, I need a nap. So, we were talking we've been talking about GPT5, right? But then our homies

come and drop a freaking banger on these uh these GPT5 guys. And so now you can see I guess we'll pose the same question we always do. Is rag dead?

No. Is rag dead? You need you have 1 million tokens of context. Maybe rag's dead. Maybe it is. But maybe it's not. You

know, I don't think so. It's not dead. It's it's not as it's just a tool. It's a tool. You you still need it,

dude. But how baller is it to drop a 1 million tokens while all these people are just talking about backlash and everything and you just keep rocking with your own making it better. But I do have to ask this question. What is that logo? It's a butthole.

But what are the hands doing? And once once I saw it, you can't unsee it. Oh my god, dude. What the You know, I just like I gotta bring it up. There's a whole joke of like every

AI logo looks like a some kind of version of a butthole and now you got two hands around the butthole. I mean, it's just and I don't want to be childish, but come on. They're not even They're I think they're leaning into it. I think they're leaning in. They probably listen to our show and they're like, "Hey, these they found out."

I mean, all right. That looks like two hands around a butthole. I don't know. I

don't know what to tell you other than that's what it looks like to me. Once you see it, you can't unsee it. So that's okay. I'll take the 1 million context win. But Claude Sonnet now sports a million tokens of context.

Uh which is great. More context. It's, you know, a little spendy. You know,

cheaper than Opus, but you know, still spendy. They highlight bolt new, which we'll talk about in future. Anyways, but yeah, that's sonnet is still my daily driver for writing code and that has not changed and I don't think I don't have plans of it changing soon. We'll see. No,

but there's more from Enthropic. So, Anthropic, you know, has always been a proponent of AI safety and so they're releasing some more things. So, Enthropic says, "Some claude models can now end harmful or abusive conversations."

So, if you're a user and you're taking it down a rabbit hole and Claude ends the conversation, you should probably ask yourself what you were doing. It's probably not a good sign, but uh essentially just more safety measures trying to probably protect people from theelves in a lot of cases. What if COD will start swatting you in

the SWAT team? That was the that was the joke in the past is that in certain situations it could alert authorities. That's true. And that that's kind of concerning behavior. I like this better. End the

conversation before it gets too far. Yeah. Don't call the SWAT team on someone. Yeah.

But there's uh so then there's some other releases. This is specifically to cloud code which I use quite a bit. And this one is really exciting to me because I have a background in like teaching. And we released our MCP

masterp, you know, is like a MCP server that basically acts as tries to get your IDE your you know your agent to act like a teacher or an instructor. But now Claude had cloud code has announced learning mode which is pretty cool because it yeah it turns the cloud code agent from I'm gonna do this for you to I'm going to teach you how this works and I don't

know how many many people will use it. I haven't yet tested it. I'm curious to test like RMCB doc server with the learning mode and see if it like really does a good job of explaining it. That's a good idea dude. We should try

that because I think it would go deeper than just doing it for you. take the content of the course and then try to teach you the content of the course. So, I'm curious. You know, this was announced last week. Haven't had a chance to test it yet,

but I do think this idea of not just having the agent do it for you, but teach you along the way is important because as we've all seen, the code generated from this these agents can still be insecure if you don't understand it, if you don't review it. It's great for demos and prototypes, but if you really want to ship something in production, your code still has to be reviewed and you people have to understand what the code's doing. Maybe

someday they won't. That day is not today and that day is not tomorrow. So the ability to have it act as a teacher, I think it's a huge unlock. Yeah, I I hope other I hope cursor adds some kind of like learning mode. I hope VS

Code has some kind of learning mode. Every uh Agentic editor, it's not that much harder, right? You just have a different agent with a different system prompt. Correct. with a different goal. And we're not saying that because we

have an MCP course that it would be beneficial, even though that's exactly why we want it. So, yeah. But it just makes sense. People want to learn stuff. Yeah. As well as have it do it for for them. I

mean, I hope people use this and I'm my fear is that no one will and they'll just like Yeah. But I I do I drastically I desperately want maybe because the terminal is not a good surface area for learning. Maybe the the IDE is better. I think the IDE would be better, which is cursor if you're

listening. You know, Windsurf, you're still out there, right? Come on. What's Windsurf dude? It's good.

Is it called Devon now or is it still Windsorf? I think it's still windsurf. I didn't mean anything by that. If you're watching, Windsurf.

Yeah, come on the show. We still love you. Come on the show. Win us back. Win us back. I can be one back. All right. Claude

code. More Claude code. It has a new model option. So just a simple little

change you know instead of just selecting opus or sonnet you can basically say like I want opus to plan it and then sonnet to execute it. That's a trend going back people doing different models for different things. Yeah you don't have to use the same model for everything. So cloud codes

releasing it. I think you know if you're building your own agent take up you know you should if you're building your own agents you should look at all of these like cursors cloud codes and what they're doing because a lot of those techniques are probably relevant to you and same thing we looked at like with what Jules is doing. Yep. If you're building agents a lot of the

techniques can be applicable for the agents that you're building, right? There are patterns that exist. Yeah. So it's like oh maybe a better

model for to plan out the the the task and then pass it off to a slightly lesser or cheaper more latency sensitive model to actually execute. All right. So that is uh the anthropic updates. I like that. That was good. So Microsoft is you know they're still

in the game. They're still here. They released something called PML prompt orchestration markup language which is cool because this be? Yeah, we've been talking about, you know, whether in Maestra, like should we

have prompt templates? Like what does that look like? Is there a standard? I don't know if this is going to be it,

but I think something like this needs to exist. Yeah. Like problem with these is like the abstraction, but I guess if there's very expressive like an XML type of structure, I just don't I don't know if this is going to take off or not. That's the that's the threat here.

Yeah. because then you have to you have to deal with all this later if it's not the standard that you want to do which is why I think most people just like writing strings and they'll just manage it themselves you know like no no no magic yeah I mean I I don't like it's it's XML so you have the role the task I guess you can put images directly into

it you can specify an output format just compiles right to the object structure yeah and then it it just compiles it right so it's not that different than like you know React right of just like compiling into HTML. I mean, that's the goal, right? Is you have some kind of structure that then compiles to what ends up being a string that becomes

your, you know, a way to structure your cont your context, right? It's all context engineering. Yep. Will this win? I don't know. I haven't looked at it enough to have a strong

opinion, but overall it kind of makes sense. Like you I think there is a use case for having prompt Yeah. templates that that can do things and you can insert variables. you can do, you know, similar things to what you

could do with, you know, markdown or HTML. I just don't think it's very novel at this point because there are so many preempt and Yep. you name it, dude. Like Yeah. So, I I think in the one case like

MCP was the right place at the right time. This feels like it might be just a little too late and it I don't know. I think it's going to take a long time before one or two standards emerge. That's the problem is that MCP

has kind of like kind of become the standard. It's going to be hard for anything to unseat MCP even if it's not perfect. I do worry that this is like there's too many other like disperate examples that the idea that Microsoft can come in and release it and it kind of feels like when Google released a 8A

although I don't think there's just one winner yet. So maybe there's a little more hope but I'm not I I would bet against it but you know. Yeah. Like the thing difference between MCP and this right like this you have a

fallback which is just writing strings with MCP you truly needed maybe you didn't I don't know but you truly needed some type of format for tool calling and getting all the tools and stuff so I guess we'll see yeah I mean I I think people will adopt and you try it and I I do think that you know it's one of those things it's not that hard to probably eject from this because you can

just eject into a string so if you if it's the lower risk. You could try it and you can move away from it. I don't think it's that big a deal. So, I hope more people try this or other types of things. You know, maybe at MSRO we'll

experiment with what templates could look like at someday. We're not we're not quite there yet. But I do think that something like this there needs to be a few standards. You know, ideally one, but

yeah, hopefully just a few that emerge and maybe this one will will eventually take over. And yeah, so preempt is was mentioned by Sherwood. Yep, that's a good call out.

Cursor invented prompt templating based off JSX. So yeah, it is one of those things that but it's whack though. But anyway, yeah, it just it didn't it hasn't landed right. And then there's GenX too. Gnx,

they also have a very similar thing. Disp. Disp. Yep. There's a whole bunch of them that that's part of the problem, right?

instructor. It It's the meme that you Microsoft said there needs to be a standard. So they introduced there's 12 standards and then they introduced a new standard and now it's like now there are 13 standards to do the same thing.

Uh will one win? We'll see. It seems like it's it's not it's a feels more of like a trivial problem than like the core problem that MCP solves. So I don't I think it's

going to be hard for one to be dominant. I think you'll have niches of people that love this because it gives them something and they'll be teams around that but I don't think it'll be become like a a massive standard that in the way that MCP has. Yeah, as long as any of these tools compile to LM model messages that you can send to the API, uh you can do

whatever the you want. So, we have a few more things and then we're going to get out of here. So, we have some news from Bolt. I don't know if this is actually news if I'm being honest, but I feel like this is just an announcement. I think it's

existed for a while. Yeah. But they announced last week Bolt Cloud.

So, it, you know, it says today this changes. I don't know if they actually if it was a change or they've slowly been building this up and releasing it. Now, they're kind of just packaging it and announcing it. That's kind of what marketing does. But you know with

Netlefi and Superbase they're basically saying that it's a unified platform for building and scaling billions of users. I think the challenge that a lot of these tools face is how do you go from toy or prototype to production. Well you need some you need a good deployment provider. You need good database and O. And so I think Bolt is

saying like with Netlefi superbase you have Bolt cloud you can now production production labeled it. Yeah, they b Yeah, they're basically saying like now it's production ready. Yeah, I still think weren't they already deploying to Nellifi? Is it like now they're doing it on like I said, I think it's this is this

is more of a marketing release in my opinion. I think all this stuff existed two weeks ago and now it's just like maybe they added a few little things and they're saying, "Okay, now it's ready to officially announce that it's here." Even though people were using Bolt, I think most of it was already there for it's been getting added on to over the

last couple months. I think maybe before like you would create Nellifi projects but then you'd have to claim them. Yeah, maybe claim them later. So now it's probably all tidy. Now it's a much more integrated experience I suppose and they're just

using Netleifi under the the hood but it feels like you're still part of Bolt which is cool because it makes it easier and same with Superbase. You don't see that it's Superbase powering it. But I think with these tools it's this something like this needs to exist. I still question the complexity of these

of apps that you know can be deployed here. I still you know really like lovable. I like bold. I like replet. We've used them you know to build like

prototypes and I think a lot of people have but I think that if you've really dealt with working with these agentic you know code editors you know that there's they still fall short and there's obviously they have security issues. So, if you don't know the code, I'm really uh really still skeptical of how quality of some of these apps can

be, but I know that there are exceptions and there Yeah. a lot of live uh projects out there that are probably insecure as hell, but they're live. Yeah. But there are um like the Bolt Hackathon had a lot of

cool projects. It's just like how many of those are like true production type things? I think that's where like they're all these companies are trying to figure out like how do we get our users into like a production use case that generates them revenue, you know? Yeah. And eventually, you know, I gota

imagine between Replet, Lovable, and Bolt, they want to uh they want to get you from just like a $10 a month plan to some kind of enterprise, right? Like almost always an upscale enterprise motion. And so they're probably thinking, if I were them, I'd be thinking, okay, well, people say that vibecoded apps aren't secure. Let's have a security product

that scans it, that tells you if it's secure or not. It can be agentic, right? There's all these different tools that you can layer on that can if you if you have enough tools later on, maybe you can get it to the point where it does feel like you can safely deploy this stuff. Yeah. at scale. They want you to they want they want the

types of projects just like Claude, right? That require lots of prompting that you're just iterating, so you're spending more money. You know, that's the type of projects that you want. If you're just viating a UI, like maybe you

exit the iteration loop at some point because the UI is done or something, you know, and if it's not a production app, you're not going to come back to it and maintain it the same way. So, you're probably losing money. But then again, I haven't used Lovable in months. I still pay 20 bucks a month. They got you. They got me, dude. They got you. geniuses.

Yeah, they still got you. Just in case I want to vibe code something just in case. Just in case. You got I'm assuming your

credits don't roll over. No. All right. Uh, so a few more things and

we'll get out of here. These are just some you occasionally like to share interesting projects that we find on GitHub that maybe you'd be interested in. So, a couple here that were shared among our team. If you're Claude Code users like we are, there's a Claude code usage monitor, which I haven't tried, but

seems pretty cool. You know, it could tell me that a couple weeks ago when I spent $100 in a weekend, which, you know, it's probably a lot more now. Yeah. That time we spent $8,000. Yeah. The time Yeah. If Tyler would have

had this that those two days he spent eight grand in usage. I guess he wasn't using clawed code, but you know, if you're looking for more detailed usage, that's pretty interesting. Give it a star.

Yeah, give it give the thing a star. And another one. So, this is just a symptom of a a problem. Yeah. Right.

And it's actually this. It's not the same user. Looked like it was the same GitHub user. Uh but this is kind of a

symptom of a broader problem that we need a ruler because our codebase is the same thing. We have like cloud code rules, we have cursor rules. I mean you get all these different rules, right?

That that are kind of end up in your repo and you have to talk to them in slightly different ways. So the idea is can we have one file that all the code editor the code agents can use whether you're using cursor cloud code because not everyone on the team uses the same tool. Yeah. And so I I do think that I don't

know if this is going to be the standard that emerges, but I hope so. I hope there is some standard that that says you can have one file and you get Yeah, it all ends up working out. I think Claude contributed, right? Is that anthropic in the contributors or Looks like it. Or is this Claude code?

Claw. Yeah. Well, yeah, code. Open hands agent.

Nice. Yeah. I mean, you got some It's got some stars. Oh, did we?

My computer died. Oh, we lost the lights. Look at that battery.

Here you go. No, actually. Oh, anyways. So, give those things a star. And also, you know, while you're there,

find the Monsterra repo. Give us a star, too. Yeah. Give us a star on GitHub. We We

appreciate all your stars. And if you want a copy of this book, you know, you want this, you can get a digital copy. Go to master.ai/book.

Get yourself a digital copy. Um, if you want a physical copy, just uh find me. I can if you're in the UK or the US, I can probably uh get you a copy. So, find

find me or Obby. And you can do that right here. Yeah. We're in SF this week or we're going to

a bunch of events. So if you need a book, you'll see us out in a boot. Yeah, we'll be out and about in San Francisco. So anything any closing words of wisdom? But before think about that,

go to YouTube, subscribe, follow us on X. You can watch us on LinkedIn. I I got got to get better about promoting all the different places you can find us. Apple Podcast, Spotify. Hey, if you've

been listening for a while, go give us a review on either Apple Podcast or Spotify because that that would be helpful. Helps more people find us. So, please do uh give us a five star review.

If you are going to give us less than five stars, just don't do it. Just don't do it. Just don't just don't just just be like, "It's not worth my time." Yeah, they're not that bad. I'm not going to give them a negative review. Don't us. Okay.

Yeah. But but if you like us and you're willing to do just two minutes of effort, go give us a review, please. We would appreciate that. We try to have fun. We try to bring you the news. We try to uh bring you cool guests like we

had today. Crazy guests today. Like honestly, the the we've had great guests, but it it feels like they're just more frequently we're just getting people that are saying like, I love the show or I want to come on. And so if if

you're building something cool, reach out. We're always looking for more guests. But we've been getting some awesome guests lately. Yeah, more to come, too. Uh, we're getting really good at I I'm speaking

for as someone who didn't do the work, but we're getting really good at YouTube. Our YouTube videos are fire. They're really educational. We released one today. So, check us out on YouTube. Yeah, Alex has been on fire.

Alex is on fire right now. So, Alex, if you're watching this, good job. Great video, dude. we have. So if you're

curious, one of the biggest questions we get at MRA, if I actually have a conversation with someone, is that's played around with it, is like when do I use a workflow or when do I use an agent? And so he did like a Mortal Kombat style agents versus workflows video and you know, not going to spoil it, but you should go watch it if you're curious on when you

should use master workflows, when you should use a master agent, and you want to figure out like how to decide, it's a good very good resource. Yeah. And overall this week we have crazy release tomorrow.

You know, shit's popping, dude. It's just we have a lot of to do. So, thanks for supporting us as always.

Yeah. Stay tuned and we will see you