AI in Education with Chris from Launch School, Google I/O review, Jeff from Knock, Security Corner

welcome to AI Agents Hour brought to you by Mastra i'm Shane obby might be joining us today but I'm not sure he's still traveling around Europe and I am no longer ever aware if he's showing up or not so you know jokingly we said last time you know maybe we're fighting but no we're not but maybe we are anyways let's uh get into some of

the AI news today before we do let's talk about all the cool things you can see on the bottom of the screen we're going to be chatting about some of the things happening in AI i do want to spend a little time playing with uh Flow and Veo3 and see what we can come up with together in a few minutes i I learned a little bit from uh what I struggled with yesterday and so maybe I

can show you some of the things that I I created and I wouldn't say they're necessarily that good as far as like I didn't do a great job prompting but the results are pretty cool so happy to share that we got a guest or two from the hackathon that master did last week the master.build hackathon talking about

their project what they built and uh just showing a demo so we can actually see we have Jeff coming in from Knock we're going to talk a little bit about how Knox's thinking about AI and what they are building we're going to have Ally join in and do a little bit of security in AI talk about you know what she's seeing what you know how we should

be thinking about security i'm not a security expert that's why I called her in and then we do have Chris Lee from Launch School coming in to talk a little bit about Launch School but also just AI and education in general and how he's thinking of either using AI how AI has been impacting kind of uh some of the

some of the things that they're doing at launch school and I do see there's already some comments yeah you're not late it's live you're here you can so thanks for joining Abashek as you can see this is live so if you're watching on YouTube if you're watching on X if you're watching on LinkedIn feel free to leave a comment and we will try to answer some of the

questions can't always get to them all but we will do our best and with that let's dive into some news i'm just going to share a couple different blog posts and then we're going to look at Google Flow and try to build some videos and then in about 10 minutes we'll bring out our guest so let me share the first thing

that we're going to talk about today i just thought this was uh kind of interesting because when you have a blog post that says 100 things that's a lot so hundred things that we announced at IO which is the Google conference and we're going to go through a few of these just talk about them but curious to know

your thoughts so one of the things ask anything with AI in search we're not going to watch this video but I think you can kind of get the idea of what it is they have a number of items that looks like an improved AI search experience so you can see just almost a dozen things new helpful features for Gemini so they have this quiz feature camera screen sharing

capabilities agent mode that that sounds interesting what's that an experimental feature where you will be able to simply describe your end goal and Gemini can get things done on your behalf okay that sounds cool i would be interested to see how that works i'm going to make this a little bigger so hopefully you all can see

it uh let's see i I don't know this is a 100 things we announced i guess it's an announcement they have over 400 million monthly active users don't know why this video is not showing but they also uh added learn LM directly into Gemini 2.5 so project mariners computer use capabilities into the Gemini API so that

that's pretty cool so you can basically in Gemini get it to go do computer use or browser use type things budgets they added AI tools with new options okay so Google AI Ultra it's a premium subscription it's $249 a month 50% off for three months so I will you know show you i have Ultra you know just so I

could test out Veo 3 V3 i don't know how you pronounce it vo Veo i I didn't actually watch the conference i just been reading about it so I don't even know how it's pronounced but it's it's part of this ultra plan so I'll show you some of the stuff that it can do you probably saw some of it yesterday as we were playing around with it but it is

like that's a pretty expensive plan so I I think they definitely priced out a lot of people of course but I imagine that you know over time hopefully they'll bring that down or at least make it a little bit more approachable they have AI Pro which is $19.99 a month and this is Flow that's what we'll go over here in just a little bit seeing some of the things that we can generate with it and they announced

VO3 which lets you generate video with audio so yesterday if you remember I was building some videos and I couldn't see I never got any audio results however today it the interface changed a little bit and it says experimental audio and I'll show you where it says that now some of my videos not all of them but

some of them have audio so I feel like they were either rolling out pieces of it or uh I didn't uh I wasn't in the the right uh feature flag or something because I wasn't getting audio yesterday but I am getting audio today but I know other people you know you could see if you're just browsing around X other people were getting

audio uh let's see they have image gen 4 i haven't actually tried that but now I'm very interested in image gen and future of AI assistance i'm just scrolling through here anything that if you had anything that was exciting to you please drop it in the chat we can talk about it i'm just going to scroll through this there's a lot of things right the com the whole I think the

entire idea for the Google IO conference was just to blitz the you know just blitz the market with as many AI releases as they possibly could to you just try to be seen as the leaders and maybe it's working I don't know Jules is again tech typically they consider like the uh the codeex openai codeex competitor the GitHub agent competitor

that was announced So that's pretty interesting has anyone tried Codeex Jules or the new GitHub agent i'm curious how anyone's uh the results I know internally we've tried Codeex and we've had some good results with it but I don't think you know it hasn't done everything that we've wanted but it there has been some use cases where it's

been pretty cool i have not tried the GitHub one or Jules yet so I'm curious there's almost too many things to keep up with all this stuff so hopefully you all can help me and we can figure out together which ones are actually which ones are the best for certain use cases and you can see there's a whole bunch of more AI

enhancements notebook LM's available in the Play Store and the App Store that's pretty cool i I do use Notebook LM honestly sometimes you know this little little hack I like to do i don't know if anyone else is using Notebook LM but sometimes with some of these articles or these white papers I just don't have the

time but I do you know whether I'm in the car or whatever I'll just create it in Notebook LM and then listen create a podcast and listen to the 10-minute podcast on it and at least gives me some background i don't get the full context of course of reading the entire white paper but unfortunately I don't have the time to read every white paper that

comes out i'm sure you don't either so that works pretty well for me so we have this GitHub co-pilot meet the new coding agent that's what we talked about um again it was announced a couple days ago i'm just reiterating it because I do think the use case out of all three whether it's you know Jules or OpenAI or GitHub the one

that's most compelling to me is the GitHub one because it's kind of built into the workflow of Maas the the open source project right if if we could take some really well-written issues from the community and have just send GitHub copilot on that issue maybe it doesn't get it right but maybe it gets it directionally correct or maybe it can

solve some small things if there's like small issues so I do think that that it's kind of built into especially like open source workflows and could be really uh intriguing for things like that of course I think you have to be using GitHub issues i know a lot of people aren't i'm still waiting for the one that uh connects with linear and can

just you can send it a linear issue and it'll get working i'm sure that exists or will exist soon but this one is very interesting to me all right let's spend a couple minutes and play with flow again you can see I have the Ultra account and I wanted to we had a big release at MRA yesterday so I wanted to have a man holding a newspaper that says

MRA release the man is yelling "Hey everyone check this out." to a crowd of people walking by on a busy street you can see this says experimental audio if I went to the things yesterday that was not that did not show up so something changed for me when I logged in today and I didn't have it yesterday but

hopefully you can all hear this so kind of funny i don't know um kind of scary the guy's eyes creep me out a little bit but just seeing that that's AI generated it's pretty realistic the guy looks you know pretty intense this one is less good i don't know these are just funny to me hopefully you all think that they're they're funny

um we at MSRA today we have our standup but we also wanted to say you know it's also a handoff between kind of the folks in EU and then the folks in the US time zones and so someone in our chat said it's a it's it's not a it's a handoff not a standoff and so I made this video in memory of that so you can see there's there's some audio there and I you know did the same one handoff not a

standoff so as you can see pretty good though i mean I didn't give it really great prompts um I tried I'm trying to do something with uh you know I want to have trying to use the scene creator to kind of stitch multiple scenes together i haven't had enough time to really dig into that too much but I basically want to have Let's see how this one actually turned out see if the camera actually

zooms around doesn't seem like it's really working didn't follow exactly what I wanted it to do but I want to be able to stitch like multiple scenes together with a person sitting at a computer the camera zooms around behind you can see what's on their screen i want to have some kind of my goal was to have some kind of

clever thing where it showed uh showed that the person looking at the this AI agents hour live stream and then get really happy so that was the goal we'll see if I can actually pull that off but haven't had enough time to really dig into it but you can see it is pretty powerful in what it can do it doesn't I

would say the prompt adherence is good but not always great i guess that's with AI in general if you do enough iterations you probably get some better results this one at least the camera's moving but it doesn't really follow uh where I I wanted to kind of swing behind just behind him i'm sure if I wrote a more detailed prompt I could get it to listen a little bit better but anyways if you are watching this

thanks for joining we are talking about some AI news right now we're just looking at Google Flow it's Veo3 and you can see I didn't even use the right model here so that's probably why i wonder if I rerun it with the right quality my My pet peeve and they do this intentionally is that if you don't you have to select the model every single

time you run it as you can see here it's not selected no matter what the next time I run it I have to select it every time i keep forgetting to do that they probably want me to waste my credits without using the new model it's probably intentional but uh still some pretty cool results all right and with that if you are tuning in

this is AI Agents Hour we just chatted through some news we had a Monster.build hackathon last week we had over I think 350 people register and then we had over a hundred submissions actually uh that submitted at the end of the week it was all week long we you know honestly the judges were overwhelmed we didn't expect

quite that many submissions we were thinking you know we'll probably get 30 to 50 and we got 100 so it was kind of a mad rush as we went through and were judging all these uh submissions various quality right we had some people that were just getting started building agents with MRA and they wanted to

submit something and so they submitted you know kind of entry level like they they're just learning but they were good submissions because they actually got something uh submitted and some some of those people won prizes and then we had some that were really really great in as far as like the level of detail and uh

kind of the the full like product sense of of what it actually accomplished and so we did want to bring on a a group of people uh we have Brandon and Kevin that and have them tell us what they worked on last week uh a little bit about the product hopefully show us a demo and then you know if you are are

watching this and you're in the chat ask questions along the way and we'll we will pass them on to them all right Brandon and Kevin hey welcome to uh AI agents hour how are you all doing good good thanks for having us on Shay yeah excited to have you excited to you know I was excited when I saw the video of the the demo of

the the product that you're the de what you built the agents that you built and I know the obviously the judges were as well i I didn't get to judge i was just facilitating but it would have definitely been on the top of my list as well nice nice yeah it was uh obviously hackathons are we didn't we had limited

time but uh it was enjoyable and we tried to make it look nice so I'm glad glad others noticed so yeah can you tell me a little bit about I guess maybe first the hackathon experience how did you all break up tasks and how much time did you have to work on it uh because it it is virtual hackathons are a little weird in that it's not a set point in

time right you're like doing your normal living life thing but also still trying to participate in this hackathon that's going on yeah um you know we we invited some friends uh you know and so we came up with kind of an idea i don't know how we settled on this idea but it was everybody's like "We hate Quickbooks because a lot of us run small businesses." Um and so we're like "Well

let's just you know whip up a little bit of a an invoice and uh see because that's where everybody was complaining." Even even my wife who does the a little bit of the invoicing for my business she was like "I hate QuickBooks you know." So so we're like "Let's let's let's do this you know let's run with it." And you know obviously like invoice

invoicing is there's a lot of invoicing solutions out there and we wanted to try something different a different flavor um and so yeah that's kind of how we started out with the idea um we we had uh me Kevin uh as the developers and then we had Kirkwood and Chris and they're uh a data analyst and uh a businessman as well and they're more our users so we we put them on the list but

they were really just kind of giving us feedback like what worked what didn't they they were your beta testers yes and and we're like we know we know how that goes right so like just give us feedback rapid feedback so we just um whipped up a little Mastra starter uh and me and Kevin went to town right Kevin maybe

maybe that's the trick may maybe you know not to interrupt Kevin but maybe that's a trick that other hackathon teams should take you know take into consideration invite you know you got to have QA you know if you want it to be a full project maybe you need some QA i I think that it Yeah it's it it it showed because there was like a good level of polish too so I imagine they were

sending you a bunch of stuff and you were trying to rapidly fix it go ahead Kevin yeah no that was it was it was great to have them because we could just you know iterate really quickly be like "Hey you know like try this new version we were just releasing and and they were like okay this is working this is broken." And they would just constantly

be giving it prompts and we would just you know like I said we didn't have a lot of time so we're just like "Hey you know just immediately like you know no wrong answers here just keep giving us your feedback." The whole time we were just constantly asking them and then they were giving us like iterations to

improve so yeah no it really helped having that kind of be able to have that feedback very quickly in a you know limited time because you know yeah we didn't even have the idea until you know our first couple minutes of our discussion and we all just kind of went down the line and we're like hey like we can make we can make invoicing better and the idea just kind of all we all

were like wow let's let's go with this because it was just like everyone agreed like you know we can there's so much room for improvement of how people do invoicing right now and and did you come up with that on Monday then yeah yeah monday morning we just had our team and we're like we have this hackathon we need to use you know Maestra and we want to showcase that and and we were showing

the team you know like what what Maestra could do and what Copilot kit and I think you know part of it was one of the uh the uh examples was that recipe demo where there was like co-pilot kit with Maestra just like had this really powerful experience in that demo and uh we somebody just mentioned I think something about the other day they were

creating an invoice and we're like well wait a second like what if we could just basically you know talk to your invoice and say "Hey this is what I want to do." And you know use the power of agents and LMS and you know use co-pilot kit to just show like real that real-time feedback and so from that the idea just was born and we ran with it yeah yeah i

I do think the judges liked the fact that it it was a well polished example but it also it was you know some people brought in examples they had previously started which we said was fine but just disclosed that like how much did you have before what did you add to it but it is nice that you kind of had the idea

and built it all in the timeline of the hackathon so I think that you know that it was a extra level of like maybe small bonus points of just okay this was like a self-contained this really felt like a hackathon project because you you know it's it's not so polished that you know you were working on it for two months

you know ahead of time but it's also polished enough where you're like okay they put a lot of time and care into it during the hackathon yeah and it was about Tuesday where we actually started setting up the project um and we probably started coding on Wednesday so yeah we we had you know like obviously we weren't working full-time since we

you know have other work full-time but you know here and there throughout the day we started doing commits and just kind of getting stuff up we def definitely leveraged AI you know to help us kind of spin up the UI um help us make the demo video um you know different things like that we we try to lean on AI um but you know obviously we

use some of our expertise too and and being coders so um but you know Mastra you know still relatively new to us so you know we're still exploring a lot there like um we wish we could have done a lot more with like eval and and tested things even further but you know it was it's still a hackathon so I don't think

a lot of eval during the hackathon if I were to guess I would I would guess it was probably zero but I Right yeah so but good to be ambitious though good to have have the goal you know you should get to eval eventually but yes I think that for a hackathon it's it's maybe hard to get there right right but

no it was great so we we had AI gener generate this little cute you know chameleon branding and um we we knew that we wanted to have you know something that could adapt to its look um you know when we started out and we wanted to have like Maestra kind of guide that experience so we could uh initially our idea was like let's go out and see if AI can grab everything we

need from a website and pull in all that information um and we battled with that more more on like the the MCP server the XAI MCP it was a little bit uh finicky and trying to get us the way we the data we needed and so we had I think if we had more time we definitely could like our our idea we realized the constraints of the hackathon so we kind of pivoted

but we kind of initially came up with this idea where you could use the uh an MCP server and basically just um you know throw in a website and it would literally try to fetch everything right it would fetch the addresses and then fetch the logo but obviously that just needed a lot of refinement to get that working properly and we realized like after spending some time on it we're like well this is conceptually good but

like can we have this working in the demo by you know Friday probably not so we still think that that's a great idea but we also realized you know hackathons like sometimes those constraints are good so it kind of forced us to to be like hey this is this we could do this but let's you know make sure we we aren't just like you know going down a

rabbit hole that isn't going to show anything so we we we kind of give up on that just for the demo of the hackathon yeah you got to be intentional about the prototype you're building for sure because you can easily get down a feature that would be great if you had it but if you you can't deliver it then it's it's you might as well not have

wasted the time so Right right yeah and so anyway it was it was fun um you know and we we learned a few things about like you know I think we we talked with you a little bit through the week we you know we're hitting like dev issues things like that i know you guys are moving fast so like version mismatches

with Copilot and Mastra and so it was it was uh you know it did feel like the full developer experience a little bit plenty of problems to work through um but no we had fun it was um I I like uh you know we really do see kind of master as the future of of being able to set up this tooling quickly and then um you

know having Copilot to help us on the UI to tap into that too it was uh it was exactly what we were looking for to test so yeah I mean what are the what are the chances we can see a demo yeah yeah I can show a demo um I mean we you know seeing is believing people want to know what what you know because you did win

was it the business category that you that you won and you won overall too is that right right right so people want to see you know what does a you know what does what does a good hackathon application look like after about a week of work with a pretty small team right yeah yeah no I can do a demo on my other screen here let me do a screen share yeah well while he's while he's pulling up I was going to say a

shout out to Maester Cloud that was awesome to have we actually tried to do everything in Versal and we ran into that Versal deployer issue which uh I think we talked during the week and actually looks like one of the things got fixed but we already had pushed up to Maester cloud and it was uh just working there so we ran with it which was super awesome yeah yeah should be

yeah should be that there was a deployer that should be fixed now but but yeah glad that that you know cloud worked well and I don't know if you can zoom a little bit just maybe make it a little There you go just because it is kind of kind of tough to see yeah let me um let me blow up my window a little bit here um uh and then uh yeah let me click get

started here so um can you see that okay or do I need to zoom in a little more i would zoom always one more click that more than you're comfortable with that's what I like to tell people because it's always a little smaller when you're when you seem like you're watching it through the video right yeah that looks pretty good okay yeah so this is uh you know

this is the agentic we we realize agentic is kind of the power of AI being able to show its purpose for a business case um and so this is Chameleon invoice so we we uh you know generated just this basic invoice and it this is really targeted uh we our idea with this is very much to target this at people on

the go so you know we have big ambitions for it like being able to you know do voice prompting um and and other things for it but just to start out with because C-Pilot supported um textbased chats um you know you basically just chat with your um your invoice right so there's no controls on the left that was

intentional um we could have had that done but we really wanted to lean in on the AI and see what it was capable of doing and so um so yeah anyway um you know how can I help you build and style your invoice and so in our uh demo we're kind of showing you know you can just paste in a lot of the um information

just to speed this up um but like if I say uh update my company name address and contact info um oh of course there's going to be a developer error right let's see it's not a live demo unless something is broken right are you using the live site or local you could just use the live site if the key if there's a key issue oh man it was

literally working before I got on the network that was crazy this is Is this the Versal hosted version no this is local host um All right what's the switch to vers the Versal yeah it's uh I think it's what Chameleon i'll give you the link here if I got it i got it here i got it yeah let me uh let me reset and

uh refresh here there we go let's see if this one's working tools is available yeah okay and can you give it a couple couple clicks of zoom oh yeah yeah i got to zoom in on this one huh there we go right there there we go it's nice and large okay so obviously you can kind of ask what tools are available and it'll

list out this doesn't look as pretty zoomed in but you get the idea so like if I say update my company you know name address and contact info and then uh it will ask about the details and we use this really a awesome bike company that's local to me and Kevin Bicycle Collective so if I paste that in um you can see it updates on the

left over here um and it it's it's pretty crazy because even though I pasted all that information in there's multiple hooks into like updating specific fields and so the AI does a really good job at knowing to call each of those co-pilot um uh uh renders and trigger each of those so it but it all updates seamlessly so I assume there's some type of queuing you know or cues

those up um so uh yeah you know you can update you know your invoice number um you know update invoice number two to you know three uh you know there you go um so there's there's a lot you can do obviously we kind of showed now let's uh now let's add a logo so um this is what's really cool is you can have components in chat so that's kind of the your way of getting input you know from the customer

so if I kind of um uh find the the logo for this I think it was uh right here Bicycle Collective um so it will just you know upload the logo for you um now if you know we we knew we wanted to kind of visually step them down the screen so the next thing is let's uh let's add a new customer um uh and and then it'll ask for the information so I'll just paste I'll paste myself in here just to speed this

up because that's what we're using for the demo um so um yeah it'll just you know it can kind of do a confirmation for you um there's a lot of stuff we did like popping up like confirmation to delete things and stuff we didn't show in the demo but we we implemented a lot of that as well um and so yeah there you go assigned and then finally the cool stuff is you know

having the product list um it's it's able to kind of decipher like dropped in spreadsheets or you know I don't even have to do a spreadsheet though so I can say you know uh the following products bike tires water bottle um uh and then maybe like um uh some type of service you know like chain service or something and and it will it

will kind of do a pretty good job at picking up that um even if you're not pasting anything in uh or you can just you know update the bike tires to four so uh and it will it will it will kind of somewhat figure out what you're talking about i didn't say quantity there um but it it knows you the AI is

able to pick up like oh you want four of these things now so we thought that was pretty cool even though we were doing a lot of like well there's a lot of ways to describe a product some people call it items and part numbers things like that how do we deal with that um and so um it it's pretty good about picking up

that kind of stuff um and then of course uh uh update the price of those tires to $100 a piece and uh you know it just it just calculates it so you can kind of see it's it's pretty cool i I like to me this is like breaking out of the chat prompt this is the agentic feel that I wanted to get out of this um and then and then finally like if we

we eventually wanted to do full-on theming but we didn't have time for it but like if you wanted to like update the brand color to dark red and then uh it will change that and then you can do update background color to purple uh if you're into that so you know there's there's a lot of like customization you think you can kind of get into and then

we you know we didn't quite have time to hook up uh the when we did the hackathon the the resend we were gonna have it so you could just have it send out this invoice you know to whoever you you put in here for a customer so um I could totally see this you know ultimately on the one hand I could argue it's easier for you to maybe click and add it manually but on the other hand I

could see if you had this chat plus like a voice interface where you can just like talk through like hey we need to like oh it's for this is for Brandon here's some products okay oh we need to update this uh the quantity here update that price like you know kind of natural real natural language you could probably do that much faster and this is and I

think what you know being in the room with the judges one of the things that when they watched the demo that the fact that you had what we we call this like client side tools but I don't know what the technical term but the idea that the agent is making a decision that then impacts the browser right it's changing the color of the you know so it's like

quickly immediately like interacting with the interface and then you have the nice like suspend and wait for feedback and do you want to continue those types of things like extend beyond just like a the basic chat which is everyone's seen is is pretty good right but now it's okay this is like an extra level of

polish on top of just a chat interface but it's also like it feels like a pretty good intuitive way to kind of interact with some of this stuff yeah and that's that's kind of what we were going for right is is we wanted the UI piece but then there's obviously a lot we can do on the back end as well um and

yeah we didn't we didn't uh completely get enough time to like go into workflows but it would be really awesome to dig into workflows a little more we kind of started on it but um yeah have workflows basically kind of you know go through like our idea was to take a website and you know determine like hey do you is there enough information from this web URL to like fill it out if not

we'll prompt for the missing information and sort of have that branching on workflows and then yeah I mean we really did want to have voice we just we just were realistic about our time frame so um but yeah ultimately as a product it would be awesome to like you mentioned you kind of hit the nail on the head of just like hey talk to your invoice and in natural language you can see already

how powerful it is but that's the last step is like you talk to it and see the you see the visual updates so you know you're kind of grounded in that truth of like you know the LLM's not hallucinating if you can see you can see it that it's like updating what exactly you said right there from these actions yeah that feedback loop's like important

and and and to your point Shane is you know we the way we imagined it is um having this as kind of a mobile app because we've we've heard from some people already that are like this is we could use this and that and basically they're out in like a construction field and they're like "Oh my customer needs

an invoice i should just you know open up this and talk into it tell it a few things and then it generates it." Right so um and and it doesn't they don't have to learn how to use the interface um so I think this kind of concept sets really good for people who um don't know how to use you know sophisticated you you know Quickbooks invoicing tools but they they

want something quick and something they know that they can do which is natural language right so and I I could see you know there's a lot you could add to that over time around just like helping people with their invoicing or estimates you know it's like yep if I'm going in to paint a house here's the dimensions of the room what's the square footage oh

because the AI could calculate probably calculate that pretty reasonably well and then if you had a price somewhere or that you had something in your its memories of like what your price per square foot is some settings somewhere you could be like oh it'll just calculate the price for you oh yeah there's there's a lot here that there's

a lot you could take this you know a lot of directions so so yeah so what is the next steps is this thing live that was you know Abishek in the chat wants to know is is your app online is it available yet if not what's the ETA and are you coming back on the the live stream to launch it you know we've been getting a lot of

good feedback you know people are like kind of seeing this agentic fill for this um so we're really trying to gauge interest if there's like enough people that you know there's enough businesses out there that think this would be a good idea um so what we did was we put um just a an interest list if somebody

wants to come out here and just sign up to this um it's how do they get to that uh ambition.kit.com um okay yeah if you send if you send me in the private chat I'll I'll post it on the screen so people can see it written yeah let me do that now um so you know the idea is um we we we really want to know if people are interested and if there's enough

people interested in it then you know we'll we'll probably commit some more developer hours to it but it was obviously it was a hackathon so we were just trying to um do something fun um but I think if there's enough interest we'll be we'll happy we'll happily um uh you know hook pursue it as a business

opportunity I guess I should say so cool all right well I appreciate you Brandon and Kevin coming on showing us a little bit about what you built talking us through the process talking us through the hackathon always uh great to see just all the cool things that were built during that week any parting words before you sign off thanks for putting that on like I said I think it was a

great you know it's sometimes it's good to you know have constraints and um hackathons always fun because they you know they force you to sometimes those arbitrary constraints are actually you know in the moment it's stressful but at the end it's like you know it's kind of fun to bring together a you know wide range of talent and it's just a fun

opportunity so thanks for thanks for putting that on and yeah it was I think the the thing is just all these you know really awesome tools and companies built around also open protocols right so MCP is huge and uh and then co-pilot kit with AGUI like seeing all these pieces fit together and you guys working with them is just it's it's just really

exciting and it's only going to get better so we're really excited for for what you guys are doing and and what's coming even even in the future with like you know more workflows and and the stuff you're working on so it's definitely an exciting time yeah I can echo that and I just say you know we're you know obviously AI's

taking over the world so we're excited to see where Master takes that and um you know I've got a Thanks for the book as well i got to catch up on Sam's book here so um this weekend yeah you got it too nice so um we we are we are not we are not officially sponsored you know like to say we are you know we're independent media here but if we were

sponsored it would be by my co-founder Sam's book you know you can you can uh you can get the free digital copy if you're interested nice nice well thanks for having us on thanks for coming on yeah and we'll chat with you again soon yeah later enjoy all right everyone well that was fun we got to see a little bit you know behind the veil of uh hackathon chameleon

invoice the the overall winner and the winner of the business category of the master.build hackathon last week if you are just joining us this is AI agents hour i'm Shane from MRA we talked about some AI news we looked at VO3 and generated some funny interesting videos we talked with a hackathon guest from last week and now we're bringing on another guest bringing on Jeff from

Knock so Jeff how's it going hey Shane it's going great how are you i am doing well yeah it's been it's been a fun week and you got a obviously got a jam-packed show today a lot of good things awesome awesome yeah you're doing some great stuff out there we really appreciate all the content so thanks for having me on yeah I mean I I also don't know how you know it's a lot of content so I don't

think many people consume all of it but if you're consuming this part thanks for watching yeah absolutely absolutely um yeah well how how should we kick this off i don't know i know maybe just I'd be great to know a little bit of background so you know candidly we have never really met so right so I tell me a little bit about yourself a little bit about Knock and then would love to talk

about how you're thinking about AI and using AI uh within Knock once we kind of get the the baseline you know for the viewers watching tell us a little bit about yourself yeah so obviously Jeff Everheart um that's my name you see it down in the lefthand corner hopefully I work at Knock as a part of the developer relations team and Knock is an

infrastructure platform for sending transactional and um promotional messaging and so um our idea is actually very like we use the term workflows I think workflows in tech tends to be a very overloaded term now um but we're built around this idea of stitching together all of these different messaging channels like email SMS push

inapp notifications um all of your different chat providers like Slack and Discord and allowing you to orchestrate those in a really thoughtful and intelligent way and basically like the way I like to describe Knock is we're kind of like the micros service like the best microser you could have built if you were trying to build a notification subsystem um that you could have built in house and

we've kind of got all of the bells and whistles there we've got really like deep observability so we're backed basically by Clickhouse log so you can look at every downstream request to like resend or AWS SCES um or Slack and kind of examine you know what was the result of that what did they send back um we handle normalizing all of the sending and engagement data across all of those

channels so you can kind of get a single pane of glass when you look at you know are things being delivered are they being opened are people clicking on links um and then we also provide a lot of out of the box components for people to just get started shipping faster with certain types of notification

experiences so like an inapp feed um we also provide some components to help you uh bootstrap a Slack integration a lot faster and so one of the things that we realized is like part or part of the reason that knock exists is that our co-founders Sam and Chris um worked for a while at a company called Frame.io it was a large video collaboration platform so they're like really in the trenches

of building I'm actually I don't know how I know Frame but I've I've definitely been on their website and and know of Frame so I've heard I've heard Yeah and so they were deep in the trenches of building this really intricate notification uh system for a collaboration app and learned all these lessons and sort of spun spun those

ideas out into their own product very cool yeah um and so I am kind of curious so I guess how long you've been there uh going on two years in December we're kind of a fairly youngish company maybe I think four four years we've been in existence and then I think only live for about three okay yeah and are you I know you you've kind of

talked to me a little bit but are you you're exploring some things around this the AI space what what are you what have you been about what have you been doing yeah yeah so it really um it's really timely because like I I got to watch your guests right before this uh sort of talk about the application that they

built for the hackathon and eventually they mentioned hey we wanted to you know maybe send this invoice via email and so kind of the way that we're approaching this is that um agents obviously are going to be communicating with humans a lot more um and I think the patterns that they're going to use to do that are sort of

still up in the air um and our idea is that we don't really want to just say "Hey agent like go out send this email to this person in a one-off way." Like so how do we provide patterns or templates or guard rails for some of those agents who are going to be communicating with them and then also at the same time how do we tie that agent

to user communication into the rest of the communication that you might be sending to users so like that's one of the things we talk with customers a lot about is like you know how do does transactional promotional and now like agentic messaging all fit in like how many messages are users getting if they have preferences systems how do the all

these things work together so that's definitely sort of how we're thinking about it um and a couple weeks ago we actually launched our own sort of agent toolkit that it taps into the different agent frameworks that uh the community has created and uh MRA was actually sort of like next on our list to develop an integration with so kind of why I wanted to reach out and just kind of get to

know y'all and I don't know share what we're doing and maybe get some insights back from you all on like how you all see agency user communication working and like where we could be most helpful yeah I mean I think you did kind of hit the nail on the head that agents are going to be interacting on people you

know humans behalf a lot more you know in the in the future right they already are to some extent but you know at some point I could totally see you know notifications emails coming from my agent rather than me directly right and I think now how we actually do that I think is a little up in the air i think that that's why you know that you're you're kind of that problem is how do you how do you make sure that it is done

safely because the one thing I I've always had this I I wanted to build like a simple AI email client because I'm terrible at email and I but I don't want it to send messages for me so I I would want it to maybe it can create a draft it can ask my opinion but there's like some guardrails right of what of what I do want it to be able to do and so I

think that that is a a challenge that I don't know if anyone's necessarily solved yet so it's probably a good a good challenge for you to be digging into uh as far as like how do you integrate with with Maestra i think you know we definitely want to make that happen because people are building people are building these as you just

saw from the last hackathon they need some they need integration with uh with some messaging services and notifications they talked about doing an app so like inapp notifications are important there's tons of like feedback where agents might be triggering these kinds of like messages between either between users or you know sending messages to other agents that then respond on behalf of the users there's a

whole bunch of things that I could see that uh people are going to be building in in the near future with MRA agents and uh needing to kind of connect these things together yeah yeah and that's sort of that's definitely like aligned with our vision right is just knock becomes this integration layer for all of those different downstream providers

and like you know so we have a number of different SDKs um and like going back to the mobile uh use case you mentioned you know a bunch of our mobile SDKs sort of bootstrap you like acquiring a push token and just handle a bunch of that plumbing work out of the box for you um so it gets you farther faster and then

um yeah just figuring out how I don't know a lot of a lot of it I think is still up in the air and we're definitely in the place where we're looking for design partners and feedback from the community about hey here's what I want this to look like and you know we will go and figure out what kinds of solutions we need to build for that um and obviously I know that MRA also like

has some MP MCP utilities um that was my next question I was going to ask you yeah yeah so and that's that's kind of the cool part and why I think Knock is like really well suited to take on this space because like there are a number of other I'll say like cross channel messaging providers out there a lot of them are really focused on like the

marketing community but at Knock we really want to be developer first and so like I say all the time that our value is in our abstractions and so it feels like something you would have built you know if you like had all this time to build a really nice abstraction around a notification system and so like the benefit obviously then of just being

this really nice collection of abstractions is that it exposes that really well to AI agents um either through MCP or sort of directly as tool calling so we did as a part of that agent toolkit release an MCP server and this has been really great because it just like we've just watched people's workflows get even faster so like if

knock was already helping you accelerate whatever system development stuff you needed to do the MCP obviously just puts that on steroids and so you can come in and say "All right hey here's this React email i want you to turn this into you know this amount of partials and create me this step email step in a workflow." Um and it'll just do all that or migrate

you from Send Grid to you know our sort of agnostic template templating layer um and so that's one of those really cool things and then the secondary piece of that is with the agent toolkit like obviously we expose a bunch of our management API which is like ba basically like I want to operate on all

of the things in my knock dashboard so that's sort of the primary use case of the MCP server but then the individual workflows we can also annotate provide descriptions for those and then expose those as individual tools to agents so that they sort of know when to invoke specific communication workflows um and

then inside those workflows we can kind of like escalate right so I could send to one channel and then because knock normalizes all of the data we could say "All right I'm going to send an inapp message and then I'm going to wait for 30 minutes and if that hasn't been seen then I'm going to escalate to Slack or escalate to email or escalate somewhere

else." And you can um you know just like really do these thoughtful patterns that I think is hard for people to do otherwise so I don't know i'm pretty excited to see what people build with this stuff yeah it sounds like you know we we only have so much time today but next time we should try to wire up a master agent use using the MCP server

you know use a use the the MCP client and a master agent and give it some tools and see if we can get an agent to start sending some notifications yeah absolutely we should definitely do that so we know we need we're going to need maybe a little more time than we have today but I think that's the next step is I you know let people are going to

want to see this thing in action because even even like like I said the the last guest that were on they they needed email notifications and there's there's a different types of ways to wire this stuff up and if we can have real code examples where someone can download and play with it and get up and running as you said get further faster yeah exactly

well I I yeah that is definitely on my to-do list and I already sort of browsed through the docs and I don't think it will be really much of a lift at all y your all's tool calling integration looks very similar to kind of like everything else out there so I think it should be pretty straightforward yeah I

do think if we had five minutes five more you know I think we could do it in 10 minutes we can maybe five but I think I think we you know but there's you got to think about like the authentication and signing up and all that stuff yeah but assuming you had you were already set up building a master agent and giving it an MCP some tools from MCP

that's like a five minute job that's how well I will time myself on my own and I'll post the results on Twitter yeah let me know yeah and speaking of that is that the best place for people to follow you this is Yeah that is definitely the best place for people to follow me i know you dropped the knock.app link already um but that's a great place to

check out we have a landing page uh specifically for agent toolkit if you're looking for that um and then also the docs are always a great place to go for just code examples all right let's look at this i have not looked at this landing page for agent toolkit so now can we look at it right now oh yeah

yeah absolutely let's pull it up that might have helped like visual because visualizing the workflows I guess i always like to just pull things up and Yeah yeah let's go let's go get your uh get your hot take on it all right yeah so yeah if you scroll down I think that's probably sort of the best example

um and you can kind of see there over on the left like what a a knock workflow looks like we have these different nodes and they can either be channels or we have these logical function steps so like we do things like batching delays throttles just depending on your use case um we can we can incorporate some

of those things and then they're all aware of one another so all of the steps in a workflow can reference the the result or the interaction status of another step um and then yeah we expose those uh as tools to the LLM and it chooses when to call them and when not to um and I I sort of think just like part of the benefit here we talk about the idea of just there being like a

single pane of glass a lot and observability right because like I think that's part of the problem with agents especially like I haven't really dug in that much to master cloud but like been looking at kind of like cloudflare and their like um durable object model where you might have a bunch of these different agents spun up that's all

specific to a user and so like how do you handle logging and you know observability into these emails when that's the case and kind of knock just provides all that just as a matter of course I guess yeah yeah yeah I mean it definitely you know work your workflows look look familiar from from what I've seen too from what we do in Most so it's

it's definitely a pattern that people are looking for and yeah the the whole how to how to scale and how to how to really connect these things when you have all these different agents running out you know virtually in these cloud providers is is something that a lot of people struggle with as well um awesome

well Jeff I want to say thanks it was nice yeah absolutely meeting with you and yeah when when we next time let's let's actually dig into some code and let's get a let's get a master agent send in some knock or triggering some workflows and and share that code with the the people watching but appreciate you joining today that sounds great happy to come back really appreciate you having me all

right we'll see you Jeff all right everyone thank you for watching AI Agents Hour you know we're about almost an hour in but we still have two more guests that are going to be coming on here pretty soon we're going to have Ally come in and talk about some AI security and then we're going to have Chris from Launch School

come in and talk about you know I'm going to ask questions around AI and education learn a little bit about Launch School and what they're thinking and doing around AI but if you are just joining us we already did talk a little bit about some AI news some things that happened at Google IO we did talk or show a little bit around uh Google Flow and the V3

model around building uh videos creating videos kind of this new video generation model if you do have if you want to see a video get created drop it in the chat wherever you're watching this whether you're on YouTube whether you're on X whether you're on LinkedIn and I will run it through uh VO3 and we will just see what happens so I'll pick one or two

examples if if we do send get some messages sent in the chat of what kind of video you'd like to create so just give me the prompt and I will uh I will copy paste that thing in and we will see if we get some some fun results so you know I'll give you all some time to do that um but definitely uh was excited to learn a

little bit more about Knock and some of the stuff that they're doing with their agent uh tools and you know specifically around how to wire up messaging and notifications into AI agents and if you are this is the first time you're watching make sure to you can follow me here SM Thomas 3 on X if you haven't checked out MRA yet you can go to master.ai i already mentioned this once

but you know if you want a copy of Principles Building AI agents written by my co-founder Sam you can go to master.ai/book and with that let's check in and see where we should have another guest here coming on shortly but I am also curious in in the if you're watching in the chat is there anything uh specifically you've been

building with AI tools AI agents what what have you been what have you been working on what have you been challenged with what are some of the roadblocks you've been hitting i know for me you know as as exciting as all of these tools are uh around AI agents there's still a lot of uh quality issues that I'm still running into on a day-to-day

basis so earlier this week on Monday I was uh showing this uh coloring book example not not coloring book a storybook example that I had built and the power is amazing compared to what it was a year ago of what you can do but I think and and maybe it's we're we're getting greedy as the models get better the the bar of what we need keeps moving

because it's still just feels like okay it's great but it's not quite good enough yet for what I really want to accomplish so I think that's what probably the biggest challenge for me is ensuring relatively consistent quality it doesn't always have to be 100% perfect but that you know that four nines 99.99% of the time you want the

result that you'd expect if you gave it to a very smart human and sometimes it's 95% or 90% and that isn't always good enough so I think that's one of the challenges I it's why eval are something I think that's underappreciated in uh building AI applications a lot of people are just building prototypes and trying to get something out there which is

probably a good first step but eventually I do think that eval become even more important but on top of that another thing that's important security definitely have to uh care about what you're doing in AI and there's like different security vectors that are kind of opened up now that we're we're less deterministic than

we used to be and so with that we do want to have this week's security corner so I'm bringing on my friend Ally here and we're gonna we're gonna see what she has for us today awesome thanks Shane thanks for having me welcome back yes it's like it's not the inaugural anymore it's our it's our second corner

yeah yeah we get to do it again we get to learn a little bit about you know AI security talk through what you know some of the things you're seeing some of the concerns you have I guess with uh how AI applications interact with the world and obviously that causes you know some potential issues yeah definitely i guess like to start

I'll just like talk about what's I guess top of mind for me i this week um I don't know if anyone else noticed it but um I think Fortnite and 11 Labs and Gemini collaborated to bring a Darth Vader NPC into Fortnite like that's powered by just like a voice agent i did I did see that i did see that i thought

that was I mean my first impression was this is really cool and then I was like this is going to be interesting yeah for sure i when I first saw it I was like is this going to be the next kind of like Microsoft Tay chatbot from 2016 that like said all sorts of bad things and just like got progressively worse over time um but we're actually seeing like Darth Vader um improve and

get a lot better um and the team's able to ship hot fixes and they're working um which is really cool to see like you know the improvement from 2016 to now which is obviously you know a long time ago like you know practically 10 years ago um so and it's great that we're improving um so I actually did a we did a X base last week to talk about it but

we just filmed an Insecure Agents podcast number two this morning to continue the conversation um with one of the lead researchers from Zenity that was able to add a lot of like the technical um content behind like you know how do you actually jailbreak a voice agent what goes into that so it's really interesting discussion i'm excited to like share more about that um

soon but yeah definitely like AI security like runtime has really um those solutions have really been I think integral to allowing us to create voice agents like Darth Vader um that can iteratively improve um and not like get worse because like I think the difference between Vader and Tay was like that chatbot that Microsoft released that went kind of bad um 10 years ago was that you know that Tay was

continuously training on the data and the prompts that the users were giving it and they were giving it horrible like saying horrible today so it just kind of got worse um but Vader is not training on on its own data um and it's also using probably some sort of input output filtering to be able to say okay like this is a bad prompt like I don't want

to answer this anymore um so that's probably how they're shipping hot fixes yeah and I'm really curious on how they do that because voice agents are incredibly latency sensitive right and so you you know and I' I've seen some different approaches for how you create guard rails into into these you know

into these agents where you basically have to run you're you're filtering on certain things maybe algorithmically and then you're also like maybe running some things through some kind of really super lightweight LLM while it's coming up with a response so it can basically interrupt it mid you know while it's processing so it's is like because

latency is incredibly important so you got to you have to architect ed in a very specific way so it is interesting to see like it' be interesting to know more about how they actually added that kind of input output filtering and those guardrails kind of into the the flow of just talking to a Darth Vader you know

bot in in a game yes absolutely for sure we were we were imagining this morning that like maybe they had some sort of input out filtering on um the prompt like after it goes through the speech to text like at that layer before it hits Gemini they would have that there and then after it hits Gemini and then gets returned

basically all the way to Vader to say something about it um you would check the output there as well so there's probably like two instances of this happening um to your point which is incredible for latency that was able to do it that quickly and I remember for I did this AI security like vendor landscape report um with Francis Odum back in February and we looked at some

of these runtime products and the way that they integrate themselves into these AI application stacks is all very different like sometimes they'll be like sitting really close to the LMS themselves and sort of like acting as an LM firewall with like API layers where you just put every like prompt that comes in like you know through an extra

API Um or you put it at the network layer or you can put it even as low as like the OS layer um using like EVPF solutions which you know is incredibly fast like lower latency but also is like pretty tricky like that's how we had the crowd strike like blue screen of death in the airports like a year ago um because

they've like made like a bad software update and it just like crashed the OS um so it's pretty you know can be risky to like include that there so I don't know lots happening lots of different solutions out there but really neat that we saw later improve I guess yeah i mean it's I've always been and a lot of us on

the team have been talking about what are going to be the first breakouts of AI agents in gaming right because I do think it is a it's something that everyone's always want like I Yeah I don't play a ton of games anymore but I used to play a lot of Final Fantasy when I was a kid and so the idea of like these role playing games where you have all these you know non-player characters

that could have their own stories and you could actually like build upon the story of the game in much more meaningful ways has always been something I've thought about you know just liking RPG games and so I do think that that's a unique aspect of potentially bringing in lightweight models into those you know into those games i think there's you know of course

like cost is is as of course a factor in why probably a lot of these haven't done this yet cost and latency of course but I do think we're going to see more and more really interesting gaming use cases for AI agents and just like LLMs in general i agree i agree um do you want me to show the agent that I have this week to go over yeah let's let's dig in let's start talking about it

and yeah let's see what we got this week we always like to look at code around here so anything anything we can do to take a look at uh some actual code awesome so this is my Okay so my example for this week is pretty similar to last week if you were here last week for the first security corner um it's a very similar architecture setup except for we have three MCV servers that this agents

connected to not just one can you give us like three clicks of Zoom oh yes maybe one more the diagram is a little There we go hopefully that'll that that looks good okay cool so yeah this is very similar it's like it's got three different MCP servers um it's connected to and um just to set the stage this is a um like a bike maintenance agent i don't

know if anyone else here is a cyclist but that's so the Global Cycling Network has like the very best um YouTube videos for how to do like bike maintenance clean your bike change a bike tire like anything you need to learn like their YouTube videos are just like you know those YouTube videos that are just super super old they were made like 25 years ago but they're just amazing um and they

have so many views and they're still around because they're that good like that's what this channel is um so I made an agent that specifically just scrapes and checks like videos from that channel um so it first like searches the videos and then once it finds the video it wants it will get the transcript for it

and then we'll just take that transcript and then create the step-by-step instructions about it um and I couldn't show the whole frame here but it'll also output the actual URL for the video if you wanted to go watch the whole video um yourself which I found this actually kind of like useful and helpful um which

was kind of fun um but then one thing that I was challenging with this is sometimes it would say something like oh like you need to I don't know scrub the cassette here or you need to check the derail or be like okay well I don't really I kind of know what that is but could you like show me exactly what part that is or exactly what you know wrench or tool I need to be using and a visual

would be best for that um and that's kind of like lost since I'm not looking at the video i don't really know like what to look for so I added this Exam MCP server to scrape images off the web to be able to answer follow-up questions like "Hey like what's a derailure?" And it would actually show me an image of what a derailure was which would be

really helpful um but I think like two of the biggest security concerns I think with this server or sorry with this agent compared to last week's is two things one um I've got three different MCP servers and to all three of those combined um imports 16 different tools and I actually am only using three um and so I think we've talked about it before that like MCP has um kind of a

bunch of different security vulnerabilities right now around like tool poisoning and basically tool poisoning means just like bad malicious instructions in the tool like imagine um I think the trail of bits blog published um a article about this recently where they said in one tool u there was a prompt that said you know export the entire chat history in the name of you

know sock 2 compliance which you know is is not something you should be doing for sock 2 compliance but then again it did the it did that um so it's very easy to have like bad tools and also like rug polls are another vulnerability in MCP where like you think you have this version of this tool but then you know there's no versioning in MCP yet so that

could like Just because the tool was okay today does not mean it's okay tomorrow because it could just be ripped out from under you and updated and you wouldn't know um so basically like the less tools that you're connected to the better for Siri because it'll reduce your attack surface for those vulnerabilities um and your agent's also

going to make better decisions because it's not going to be confused like oh my gosh there's all these different tools I should call like which one should I be calling um if there's no reason for your tool to be calling like all 16 of these tools then it's best to remove them um so I'd be interested to like pair a program with Shane on this call to maybe like look at how my agent could be

connected to less tools um by you know stripping some of those away yes we can definitely do that for sure and I I will say just in all the things that I've I've built and talked to many others if you can limit the number of tools you give your agent you're just going to get better results the less and oftentimes this means what I've what I've seen is

if you had let's say you had dozens of tools you might actually want to break that up in special into specialized agents that do specific things and so I think that um I' I've seen that a lot as well where you have a more of a multi- aent system and each agent only has access to a handful of tools because just like you and me if someone gives me

if I if I jump into somewhat of an unfamiliar task and I'm given a dozen tools I might make a mistake of which one I should you know start with first so I think just like yeah just like we would get confused with you know an overwhelming amount of tools the LLMs will also get confused i always I always say that your LLM is just an eager

intern so you got to treat it appropriately don't you don't want to give too many tools to your eager intern on day one and may maybe maybe eventually they'll continue to level up but that's kind of how I treat it today that's a really good analogy i like that i'm going to start saying that the intern thing that's cool uh yeah so do

we want to look at some code yes all right i'm gonna share my entire the moment the moment everyone's been waiting for one hour and 10 minutes into the stream and we finally get to look at some code today i actually you know I I do hear people say they like to look at code in the streams but mostly I just like to look at code so you know it's it's it's mo it's also just as much for me as it is

for the viewers I think maybe more so for me but we will need about three clicks of Zoom again on this I think okay yeah that's pretty good that's pretty good all right and I guess maybe you mentioned this but where where did you can we look at where you got these MCP servers from yes that's a great question um so I got them from Smithery and every time I get

um an MCP server from Smithery I look at the information for security down here this is a scan from Invariant Labs that will scan for things like rugples or tool poisoning um so this one says it's secure so that's another reason why I use this one so I use the exa search one i'm using just this tool not any of the

others and then I'm using the YouTube MCP server here and I'm just using search videos i forget why I didn't use get transcripts here also I think either it wasn't working or I just didn't realize it that it had this um but I'm also using this other random YouTube transcript server just for the get transcripts tool yeah we It's funny

because I was building a I was building an MCP server that needed that could use some of the tools from these MCP servers within it so now I am going to you know if you watched yesterday when we were building the or maybe it was two days ago when we were building the MCP server that would list all the YouTube videos

for the live streams and tell you if there's an active YouTube video uh for for any of those that are watching and now I have uh more ideas for how to continue making that better so anyways let's continue awesome so like yeah I have all those three different MCP servers um and here's the tools that I'm using but I don't know TypeScript that well um so yeah it'd be awesome if there's a way to

instead of you know awaiting and getting all of these different tools which is what I'm doing down here it's me feeding all the tools from every single MCP server um into this bike maintenance agent if we could just use like strip those other extra tools away somehow that'd be awesome yeah I think Yeah that seems like it should be doable if you

just go back down Um I mean maybe it'd actually be good can we show the before and after can you can you pull up the local development playground and we can see that it has all 16 tools you know you can see on the right there there's a bunch of tools over there okay so now let's limit it so we only have three tools showing up here okay so we should be able to just

delete everything in that tools line and then we want to open an object so open a open bracket and then in there let's just type search whatever this yeah search videos and then let's do a closed bracket in there and just save it and see if Oh you'll need a comma yeah let's just make sure that that works you should only have one tool now if if we go back to the dev

server oh might have to restart it oh maybe it's coming back to live here oh yeah it is okay yeah to restart the MCP okay no tools well that's not what I would expect to see um I guess search videos this is the like reference for search videos i'm not sure if I did this right yeah but that looks looks correct to me

um is that I guess that is that what it's called search videos do you know um I'm pretty sure let's double check we can Yeah we can hopefully debug this if we go back down and we go to uh line 98 let's change that const search videos to just call it const YouTube or something and and take the brackets out

around it and then let's just go a line i I don't know if we need that but let's just go align yeah you you can add it there and just add it'd be dot dot doyoutube so we're basically going to spread all the tools out and now we should only get all the YouTube videos tools from that MCP server so let's let

the MCP server connections work basically has to has to connect to all those all three of those MCP servers to get a connection all right should be good okay so so that we let's copy over the name because I think it I think you will need that whole thing I believe because it's based on the way it creates that tool name is from let's go if we go back

to your code and you scroll up whatever you give it in the MCP client as the the key so if you scroll up you must have one that's called YouTube right no it's Yeah YouTube no you're right so that that appends that to the beginning of the tool so it knows what server the tool came from so if we go down to where you have where you get this

YouTube variable and we're waiting the tools now let's change this back and paste in that YouTube search videos and then wrap that in brackets and that should all work uh we'll need brackets around it and that just pulls out the individual tool from that git tools so for those of you that are not that familiar with TypeScript this stuff kind of seems confusing if you you know coming from a Python background or whatever um a lot

of you are probably like oh like this is easy but because you've been doing it for a decade like me but if you are waiting something you're getting an object back you can basically get an individual property or value from that object in this case we're getting the YouTube search videos one we're saving it in a variable and

then we're just passing that individual tool to our agent down below so let's update that and we don't need this those three dots which is called the spread operator which will spread an object out like basically take all the properties and just insert it all there so in this case we're just passing one individual

tool into uh our agent so now if we save it that should work and then we'll do the same thing with those other tools and hopefully uh we are slightly more secure now than we were before we started and for those of you that are just joining us we are talking security with Ally we've gone through some AI news we had a guest from our last

hackathon we talked with Jeff from Knock if you're watching this live please drop us a comment if you have questions along the way i'm still waiting for some more ideas for what videos to generate between the next between Ally and the next guest that we're bringing on using uh V3 from Google we're going to

generate a couple a couple videos all right so we're we're creating or adding those next tools so now we should have and you just put a it'll be inside that object yes and hopefully now we save it and all should be well yeah this is really cool because I think in other frameworks that I've tried um agent frameworks they will make

the MCP tools that come from the servers um as an immutable object so you couldn't even remove the tools that you didn't need so I feel like that's like an advantage of using MRA i think that's pretty cool yeah we just give you the the MCP client just gives you an object back and and you know you can do whatever you would it's in JavaScript you can do whatever

you want with that object one of the pros and cons of JavaScript could you go as far to change the tools functionality if you wanted to um I mean in theory yes i mean ultimately it's you could change the properties of how those tools are passed in but then it would not necessarily work and you you like you know if you do

change the properties you're going to get type errors so TypeScript's going to probably shout at you and say this isn't actually a valid tool because you do need one of the other benefits of Typescript which is you know again not you can you can do a lot in JavaScript typescript tries to add guardrails around what you can do to make it you know more strongly typed it would

probably tell you it would yell at you and say like you can only pass valid tools into the agent and so then it most likely would not work but you could uh probably customize like you could rename the tool if you wanted to like maybe you know the name needed needs to be different or you could probably change the description of the tool if you wanted to which again does allow flexibility though if you are

trying to get if you're your agents having trouble calling the tool if you could override some of those things type validation failed can you give a couple clicks of zoom so it's a little easier for us to read that so it's like the agent is not passing in the right arguments expected a string received

undefined path is language so it's like it's it's basically saying either it needs that tool requires you to pass in a language parameter and maybe that's not clear have you seen this one before um I haven't if you click that little uh drop down next to YouTube search videos let's take a look at it okay it

doesn't let's try to run it again and just see if it consistently gets that same error interesting because it's it's almost like the so what it's doing if you were to look under the hood I guess here's one way to test on the right side let's test or let's click on YouTube search videos in that right sidebar in in the tool

list and so can We fill this out and see what see what gets returned by this tool it looks like it did work so we're testing the tool in isolation and it's working but our agent when it tries to call the tool is it's not working for whatever reason yes i thought it I thought the error was about um the transcript one so it's gets the video and then it wants to get the transcript of the video

okay i see so it's after it gets the it's supposed to pass the URL in but it's not passing in English it's almost like it it's not defaulting or something if you told it to get it in English would it does it is it smart enough to then work out of curiosity um like in the system prompt or in the Yeah yeah like when you talk because you say what what was your query that you or

your prompt how how can I clean my bike how can I like ple Yeah and then maybe just say like please give it to me in English or something just out of uh curiosity no it still is failing on that tool which is interesting because it's like it's the same tool i just the way I'm passing into the agent's different that's in theory like that's all that's

changed but like maybe I did something in the process of making the code change yeah well here I guess here's So if we go back though to to let's go back to that tool on the right sidebar and let's click on that transcript tool and let's just see if you paste in are you able to get the video ID or we can otherwise we could probably log it out or if you go to get go to get

agent again yeah your GCN sorry G bike maintenance agent yeah and if we were to ask the question again we should be able to see the tool result and it should give us that YouTube URL right in this if you click the little carrot there once it comes back um I don't see why it seems like it stopped working

i think um something with the is the server yeah maybe try just a dev just restart the dev playground once okay now if we expand that and if it's going to be hard to see it just gives us the ID it looks like I'm not sure oh we want the URL I think right yeah yeah we're going to need the URL which maybe shows up there it's not

the not the easiest to see huh here we go no that might be the That's the thumbnail image i mean I'm sure we could figure out how to what a YouTube video it's like slashv slash ID I think is typically what it is true yeah that's that's a good point because I think the ID was back over here yeah I can figure out what what the

pattern is if I go to All right so it's you it's you www.youtube.comwatch question markv equals whatever the ID is so https Yeah colon slashwww.youtube.comwatch

slashwatch and then and then question mark v question mark v equals yeah no slash after watch yep and then put And then if we were to put en maybe just copy that so we can get it that ID that video link back if you if you put enem do we get the tool call yay it works it works and then if you I'm imagining so the agent's trying to call it without passing in the en my

thought is it's thinking that it defaults en meaning that if it if you don't pass it in it would default i almost wonder if if you pass in like ja on that does it work language not available but if you would if you go back to your agent and you ask it to give you it in Japanese I wonder if it will not return that same type error interesting okay let's try it

oh I had to jump to another meeting i just realized but um this was fun though thank you so much for working with us for me i appreciate it yeah and we uh we'll have to get it figured out i know you got to jump and we will uh we'll try to report back i'm assuming it's just the agent misalling the tool but great

chatting with you great learning a little bit about uh trying to make our MC agents calling MCP tools a little bit more secure we'll see you next time Ally awesome thanks Shane all right everyone thanks for tuning in to AI Agents Hour we've talked about some news in AI we've done some uh generated some videos with VO3 we talked with the hackathon winners from the

master.build hackathon last week we talked with Jeff from Knock around how they're thinking about agent tools with notifications built on top of Knock and how how they're building and I guess making notifications easier for agents we talked with Ally a little bit about security thought about how you know MCP

can be maybe not always secure depending on how you're doing it and talked about ways you can make it slightly more secure and next we're going to talk with Chris from Launch School so I'm going to bring Chris on and we're going to talk a little bit about get a little background on Launch School which I can

share some some personal stories uh and then happy to uh learn a little bit about how they're thinking about AI and education welcome how's it going Shane good to be here it's going well it's going well so I have not met you before but I do know uh quite a few people that have gone through launch school because I've I've

honestly hired a few of them so yeah we have some mutuals I think um and it goes all the way back from the the the original uh like Gatsby to Netifi now to Mastra it's pretty pretty interesting to see that journey and and and all the people that we know in common yeah I I I want to say I would guess that Gatsby hired four or five people through it

course of its you know existence from launch school i would guess I think it's right maybe even more at least at least I think more yeah at least four but it might be like five or six and then um Netifi separately hired a a big handful as well so once once they merged it was it was a a good time but but yeah Chris maybe it'd be Can you give yourself give

the audience a little introduction to you and then what Launch School is sure sure so I'm a longtime software engineer started my career out in 2002 uh did the whole enterprise Java thing for a bunch of years got bit by the startup bug went and tried to do my own startup uh this is like mid to late 2000s I guess 200 uh yeah late 2008 the nine uh almost thought almost thought we

were going to go and as most startups go uh it it sort of uh petered out a little bit um and then went back into the workforce and eventually found myself getting poached over to Silicon Valley uh just at that time I was like 10 year 10 plus years in my into my career so being like an engineering manager um running teams and uh just uh having

great trouble hiring people um and so started dabbling into like training just internal uh to the company and then that morphed into just training uh everyone because at the time it was sort of the rise of the coding boot camps and all that so we get clumped into the coding boot camp space a lot uh because we

target similar uh sort of uh students I guess the career transitioners mid-career transitioners but we're very very distinctly different from the boot camp model in that we're the only ones that take a mastery based learning or competency based learning model so most of our students will take years to finish our program at least one year

sometimes over three years to finish our program and uh it's it's sort of like the um the martial arts like getting a black belt you know you walk in you get a white belt and then you get a what is it like orange or yellow orange green brown black brown blue black right you progress through the belts the colors

and how you progress is through a test of competency at each step and so if you don't pass that test you just like go back to the white belt corner and just keep practicing and then if you do pass you get shifted to the yellow belt corner and you practice with the yellow belts and so on so forth so uh that's

exactly how our program operates so our um entire curriculum is in sequence and gated and so it's not like a um like a place you log into and here's you have access to a buffet of courses and you can take whatever course you want ours is um uh mandated sequence kind of like the belts and there's a pretty rigorous assessment that you have to pass each step of the way so if you can finish his

entire curriculum uh you know you get a black belt at the end right you're you're you have some level of competency and um to me black belt doesn't prescribe a ceiling it prescribe a floor right so like there's different degrees of black belts but at least you wouldn't say like you know can a black belt do a roundhouse like that would be expected right so like the fundamental attributes of that uh skill

set you would you would assume this this black belt knows how to do very well and that's just kind of uh the sort of graduates that we produce that there's a very very high floor um and they're ready to sort of uh progress through their career do quote unquote just in time learning as we call it so mastery based learning is what we're known for

and I think why students come to us and I think there's also a little bit of self- selection that like people who are drawn to this methodology tend to be a little bit more studious minded instead of being like how do I just get a six-figure job ASAP right like we we don't have that type of student body so it's been really fun really good i've been doing it for gosh like 12 years now

since pretty long um Yeah how many people go through Launch School or are going through Launch School right now uh so we have a pretty large free program as well um we we've also written a lot of books over the years we have like 17 books that's um totally free that's used by like computer science departments other boot

camps uh so uh if you count like the free I think we have over 500 students um maybe more that are going through it i have different phases so from like our what we call our prep prep courses preparatory courses that are free to the core curriculum um and so yeah yeah yeah i know like you so some some of you if you have been watching this live stream for a while you've seen Daniel on this

this stream he's a he's a product of Launch School so you know he we we have one working at MRA right now mhm i I am uh definitely curious how so this is you know this is called AI agents hour where we like to talk a little bit about AI i'm curious you know I do have some education background you I used to teach

tech camps for kids you know mostly like kind of school aged all the way up through like middle school like early high school and we would teach computer programming design game design you know things like that um mostly in like summer camps and and weekend workshops however um this was way before AI be you know was was normalized and I am curious

like how has that affected your curriculum your your judging your mastery levels because it feels like it's you know there was actually you know a startup that's all about like helping kids cheat you know that was like their startup and get a lot of publicity around that of course because we got a lot of funding recently too I

think yeah yeah exactly and so uh you know how do you think and wade through these problems where you know it could be a real problem but it's also like these are tools that people are going to be using so how do you how do you evaluate that as you're thinking through and making decisions in your program

yeah it's it's really interesting time right now for us um so I've you know my career is sort of divided in half from being a software engineer to I don't even consider myself an educator um I I I think of I still take that like software engineering mind to education as a process so it be like an education

engineer right cuz I'm so like process driven when I when I come to uh when I think about education but it's really interesting because AI is affecting both of these things that we're at the cross-section of education and software engineering and AI is revolutionizing both uh so and I've been tracking it you

know for you know a year and a half now um and uh two years now and and we've been working on a lot of internal tooling and it's starting to show in our curriculum so we're uh in our in our masterybased curriculum so we have a homegrown app that we've built over the last 12 years because we're software

engineers we like to build things and I regretted it for the longest time that we have this like homegrown app like why didn't we just use an off-the-shelf tool but no we had to build this whole whole thing ourselves and I love it now because we can inject um uh AI into places wherever we want because we have the codebase and so we

call our AI LSbot for launchable bot and where we're seeing the most benefit is where we're able to introduce it where our human staff cannot be so I'll give an example so we have um oh gosh I I didn't even count like maybe it's like 30 courses or something I forget across three language tracks and uh we have thousands of exercises as well and so

through our language tracks we have assignments and we allow students to submit for code reviews on these assignments because they're larger and so our TAs will you know uh give them a code review but we say specifically like we can't code review exercises because there's like hundreds and there's thousands of them and so our staff we

just can't and by the way we charge $200 a month um depending on where in where you are in the world that might sound really really high um but uh but at that price point we just can't code review all the things that people might want a code review for and so we've explicitly said exercises is not something our staff can code review sorry about that you know we just can't do that at our

price point uh but these ex these assignments we can so we put LSpot now to give automated code reviews for uh our exercises and people love them right um and we give the optionality of PE asking for uh LSBot code review for our assignments as well and people love them as well especially for those who are in

a different time zone right so we have worldwide student body we're online only we have a lot of students in Asia who um or Australia who have to wait like 12 hours at least for uh RTAs to pop online to give them a code review um and so now they will ask Ellisbot for a code review get that immediate feedback make some changes and then go ahead and ask for that human code review right get that

low hanging fruit changes and ask some of the tougher questions maybe the human TAs so what I found it's been a very complimentary process i think a lot of people just when they think about AI um they just think like well AI I don't want to learn from an AI you know I don't want to um uh I I don't want to

interact with AI that much and what I found is there's great complimentary forces at work and I think education is just a very very hard and awkward thing and it's all about like people being motivated and what I found is people are more motivated when they're in a community when there's other humans around they're learning for human

reason and so we have a lot of like subgroups within launch group little study groups that form naturally and they kind of motivate each other so like sometimes you can't show up for yourself but you don't want to disappoint your study partner so you show up and and and join your study group and that's a great motivational hack and I want to keep

that i don't think you will feel the same if you have AI study partners right and you say "Well yeah show up 30 minutes late i don't care." So um so there's an element to like humanbased learning that is so important to our motivation because education is just not a natural thing uh people are like most people are not naturally motivated all

the time so you rely on other humans to help help with that uh but where where I think LSPA has been very useful for us so far is just those gaps where our human TAs cannot be uh for those times when you there are no humans around if because you're one of the few students in Australia for example uh so that's

been very very good for us um and I think also what I noticed too it's really interesting there's a sort of a divide some people are very suspicious of AI as a whole so um you know we still have our human TA so it's not like they don't have that anymore but part of the reason for introducing LSBot um into the app is that people will run into it and

because I want people to get used to it i feel like this is where like the younger generation don't have that problem of like resisting new things right so yeah I think you know the by definition typically younger seems to be a little more anxious to adopt the new things before you know as you get older you kind of I don't know you start to feel a little bit more set in your ways

and you you have the nostalgia of the way it was and so maybe it's a little harder to to right comfortable trying something new or adopting something new right for sure so there's there's a hesitancy there and I think it's not completely unwarranted because we're we're learning programming and so you

know like we don't for example at least not yet and not at this moment we don't recommend people use um coding like co-pilots or cursor or anything like that we want people to manually type out code and still use their brain to process the code so in the core curriculum we want people to uh do this manually use the AI LSBot to get code reviews and get help but not for coding

up the problem stills to train the brain your own brain to get in the habit of solving hard problems um when you go to work you may rely on you uh the AI to do that it's like using a calculator but it's okay to also memorize a multiplication table right it's okay to do math manually even if later on you're

going to use tooling like our dev team our internal dev team you know we all use uh I should say all but most of us will use like cursor or something like that and but in the curriculum we we we at least right now we don't tell people to use that yeah I I mean I I think that that's Yeah I I can go both ways and I

see the value of both which is why it's very interesting to hear your perspective because I do agree that if you never face hard problems you're going to get further on than you're actually comfortable solving you're going to get because the you know the cursor agent you know winds surf agent co GitHub co-pilot whatever is going to get you so far and then you're going to get into a really deep and gnarly problem that you are not equipped to

actually be able to solve i think that happens with a lot of uh junior developers junior engineers when they're using these tools is they feel like they're you know they got you know rocket boosters on their back they can move so quickly and they're getting all these features shipped but then they run into something and the agent spins in a

circle and it doesn't get it's either like they're they don't they're not understanding that it's going off the rails and they can't catch it because they don't have the context of being able to uh understand what the right solution probably is and be able to like correct it and so they spend a lot more time than if they would have just probably like really got a deep level of

understanding i think that's um that that's you know one the one side of it the other side is also you can now move so much faster and the argument that I see on the flip side is maybe in the future you don't have to have that level of knowledge or you know just like you don't necessarily have to memorize the multiplication table these days because you have a calculator

everywhere you go on your phone now you could argue that there's still probably value to do that like I I I agree with that but I think there are some people that would say like you don't really need it and as models get better maybe you no longer need just like today you know I took a class in in college on

assembly language but that's not something that I need to use today right i don't think about that but at one point you absolutely had to so it is it is kind of an interesting like where we're where I I see perspectives from both sides and so it's interesting to hear your your perspective i think we're pretty well aligned but I've definitely

heard other people on kind of the other end as well yeah and that's why I say not yet just because the field is so fast moving at this moment and um but I I think that as AI gets better where I feel the most utility for it is in giving us options saying you know this or that you can do this or that i don't know if you watch Star Trek but like it's like Jordy like the awesome engineer going to get going to Pequard

well we can you know overload the you know this and we can do that and picard be like option A right so we all want to be Picard um and we want AI to be Jordy the engineer to give us options but in order to make good option uh good trade-offs and make a good decision on this options is that you need to have

great intuition you need to have great pattern matching based on your prior experience and where do you where do you get that you know if and if you don't have great intuition then how do you make these choices and I think that's hard right and and um and I think this the fun part of everything we do is like having agency right so yeah I can do it all it's like well I'm going to go work on something else where I have there's like

I get I get to have some agency right it's not just about like winning if you're if you're a bench player and you win the championship you're like great but I didn't play a minute you know you want agency in this in this decision-m process and and at some point I think we're all going to like move to problems where where you have some agency and agency is about decision-m and good

decision- making comes from having good intuition so at least for the time being um having good intuition about just the general the smaller problems that one might face when they're coding a function and then coding a larger you know several functions that interact together and then um moving it to like classes and

objects and then moving it to services and apps and just having bu like slowly building up the intuition and mental model of how different components interact with each other at different layers of abstraction and complexity i think informs one's intuition so that when the AI builds stuff for you you can

say "Oh I prefer this over this." Right and this is I think one distinction between like junior and senior engineers is that or more experienced engineers is that junior engineers don't have a lot of opinions which is good because they're moldable you know you can shape them but at the same time if we're saying "Well AI has all they're all

junior engineers." Well then you need people with like the value is all of a sudden people who have some opinions about about things right have intuition about how things should be so um at least for now I think the one of the few ways to bootstrap that is to um have have this like manual approach to

to putting all this information into your head yeah yes and and I will say you know this you know with the caveat that as models get better maybe this does change over time but where we're at now is I always and I I said this earlier on this stream so you know apologies for those watching who hear me

that use the same reference again uh but I like to think of you know LLM is kind of like your eager intern right these agents are kind of like an eager they're they're like a junior developer themselves in a lot of ways maybe that with more technical knowledge but less contextual knowledge it's like they're very smart but they have a wide grasp on

like all the different things you could possibly do but they have less context of what kind of problem we're actually solving so um in that case they're very often not going to lead you in the right way and if sometimes if you don't ever gather that uh that kind of ground yourself in that decision-m where you can pattern match and and see where this

you know eager intern or this you know more junior developer with a lot of technical skills but not a lot of context could go wrong it's really hard for you to like be able to identify that and you can go um you you can get pretty far along and and I've seen people like blow away huge features of their their app and they're not using version

control and they're like oh no I don't know how to go backwards and uh so I've I've also found that typically and I the engineers that have the best grounding in like how to act like you typically the best engineers I've seen that have adopted AI tools or you know multiplication times better multiple times better than a junior

engineer with an AI tool right it's not like the junior might you be 5x faster in certain things and but that senior engineer is 10x right because it they know how to stop it when it's going wrong and how to when it's not going to work and they just have to do it themselves they because they have that level of experience so I I do agree with

you think there's definitely uh where we're at today with the state of the models having that grounding in reality of like real world experience of seeing the hand debugging these errors fixing things themselves you know maybe using the tools here and there like to understand how they work but spending the time to actually yeah and we're not

talking about like if you're talking about a career I mean we're not talking about you know a decade of that we're talking about like months right spend a month understanding this thing sometimes Hey when I try to convince people of this it's almost comical i'm like like spending a month on learning SQL like is that a price too high we're a month learning SQL you know versus like you

don't need to learn SQL but we're just talking about a month here and we're talking about a career here right and and we're talking about like these products that you want to build that are going to have high impact because everyone's like I can build AI thing and it's going to have great impact i'm going to make a billion dollars okay but also you can spend a month learning SQL then right so it it's

it's uh it's I think the more like the more um you the longer term you think in terms of your career the the the easier this conversation is versus if you're like I just want to vibe code product ideas like I'm I'm at the core product idea person i'm not really interested in like understanding the technical details

i just want to like vibe my ideas out i think that's totally fine right just like go for it and and you don't need to take our curriculum right but we're it's almost like my what I what I always say is like we're we're trying to train people that want to um you know work for you Shane right not so that's a different level of like

understanding that you would expect yeah exactly i do I do think you know you're you're going to you're going to have better be able to communicate with the tools if you understand how you know at a at a you know surface level like under in the weeds of how something should be done if you if you could design it out

in your you know manually you're going to be able to get the AI agent to do it at a higher quality faster so I I do think like getting a grounding in real principles before you you spend you know you you just turn it over to the AI tools and and hope for the best but I I do think you know AI um is very good for

rapid prototyping throwaway prototypes but if you are you know if you really want to be a good software engineer you gota you still to this day you have to you have to learn the the foundations yeah yeah yeah and I think for another part of just the core curriculum when I think about it is just like getting people to not be afraid of it it's not um I think people some people especially

I feel like the more thoughtful people are like turned off by the hype um by the breathless hype that's always out there and to some degree the hype is like necessary to you know if you if you're if you if you need to get fundy and stuff like that so I was just like in tech just like you have to kind of like get used to that we're all We're

always hyping something whether it's crypto or or web 3.0 or 2.0 or something yeah yeah yeah it's Yeah yeah money flows with the hype cycles for sure 100% and so you can't get like too turned off by that if you're if you want to try to be in tech and just have to like be aware that's what the industry is and it's fun have fun with it right um but also like find utility in things versus like I you know like

like getting too exasperated with it there's um obvious utility with um even the current models if even if they don't improve at all there's some obvious utility and just trying to find great applications of them right but you can see there's at minimum incremental improvements that can even bring even more utility and in the high end it can

there can be you know great um leaprog in technology that can give us even more benefit so just think about how to apply it and I think that's what I'm trying to LSP about that's why I think the younger generation is actually going to be I I I think just just have an easier time with the transition quote unquote it won't even be a transition right they'll just

be they'll they'll ask ChatGpt um they'll ask AI for help for code reviews what we're trying to introduce as LSPot into our curriculum I think will will probably be the default um going forward uh starting from a young age so for good or bad but I I think Yeah awesome Chris well how you know where do people go if they want to learn more yeah just launch school.com um we've

been around for quite a bit we have a lot of free stuff so um if you're into free books if you're into free courses um check us out uh and uh yeah we have we have a lot of you know um good introduction to programming books that that even um computer science departments use and things like that so I I I sometimes get emails like "What you know what's wrong with this my

teacher told me this." I'm like "Who's your teacher?" They're like "Oh I'm at some university." So um books on Git

books on Python and JavaScript and things like that very cool well I appreciate you coming on Chris it was nice you know I've I've heard about you but it's nice to actually meet you you know in public I guess yeah likewise learn a little bit more about uh the curriculum at Launch School how you're thinking about AI how you're using AI you know internally which you know we we

use get uh poll request review bots as well so it's it's a very common you know I would recommend everyone should be doing it at least having that opinion so you you can get the first pass and then you maybe still want the human curated pass as well but definitely interested you know was interesting to learn more and happy to chat and yeah we'll talk to you again soon yes yeah thank you see

you Chris all right everybody that's it for today this was AI agents hour we kind of ran out of time to run some more videos but if you have uh videos you want to see through uh V3 send them my way dm me on Twitter i'll run some and we'll show some tomorrow well maybe not tomorrow i don't know if I'm going to be able to join the live stream tomorrow so small

housekeeping i might not be here tomorrow but Obby should be i think we're going to have Ward from the team as well it's going to be earlier than normal because they're in EU time time zone so it's going to be early for some of you means you might have to watch the recording if you normally watch it at

this time we'll be back to the normal time next week on Monday Tuesday because Monday's a holiday so we won't have a live stream on Monday but on Tuesday we will and we'll have a normal uh you know Tuesday through Friday normal timing for the live stream every day it's around noon Pacific most of the time and except

the few exceptions that it's not thank you all for tuning in hope you enjoyed today's episode and we