Multi-agent systems, edge functions support, security corner, and AI News

going to have some time between guests we're going to talk about your your travels what you learned what you did definitely there were We had some streams while I was gone uh yeah yeah you did yeah we we we kept it we kept it going but it was it was didn't feel like home this feels like home the original show yeah we but we are back so thanks

for everyone for tuning in this is AI Agents Hour i'm Shane this is Obby and yeah we're we're we're basically back it'll be even you know I'll be in San Francisco this week so we I don't think know if we're going to have a live stream in person or not but we're going to try we'll see but we do have some AI

news today we have Leonard from Hayes Labs coming in we're going to talk about multi-agent systems you know should you do it should you not what are the pros and cons what are the what are some of the other players what are they saying about multi-agent systems we're gonna have security corner with Alli and then we're going to talk just some master updates and even have a guest from the

master community come in and just show a demo of what uh of what was built so pretty cool show today yeah I'm stoked we haven't done news well done any news in a while so it was purely just pair programming on stream so this is gonna be dope yeah we got a lot and there's a lot to cover i I think the first thing that would be interesting to talk about this is kind of a a wild card i just saw a

tweet about it but there is an if for those of you that didn't hear about this there's AI startup school which is going on right now yep it's YC is hosting it and it seems like they had Sam Ultman today as a speaker and two things got announced that I thought was uh pretty cool so the first thing was

that I chat GBT is supposed to have MCP support i think it's supposed to happen today or very soon so that's pretty cool and also they mentioned that they open and I think they've you know teased at this a bit but OpenAI I think is I don't know if this is the first time it's officially been confirmed or if it's just you know more news on when it's going to actually happen but they're

going to have an they're going to release an open source model so nice you know in the the time frame that I heard from someone again I wasn't watching you know I wasn't there but been seeing what was popping up about it on on Twitter on X it sounds like maybe this month you know or maybe within a month so relatively soon we'll see maybe

it was all you know maybe it's all fake news but we will we will find out uh soon enough but two kind of cool updates I think from um coming out maybe from OpenAI yeah and it's interesting they did it at the startup school um there's so many people there uh if you look at the images that they're posting online yeah i was I mean I thought it'd be big but damn the production value is crazy too

it's better than our demo day at YC i know we I feel like we we got we got screwed a little bit but no it was it's really cool to see a lot of interest obviously a lot of people you know younger crowd I think as well that got invited to this so it's really cool to see you know all the the college students and high school kids or

whatever like whatever all ages I imagine like Yeah but a lot of interest in AI right and I think you know if you're watching this you're probably pretty interested in AI too yeah yeah definitely they It was an application process for it so not everyone got to go so I'm sure there's some some wild builders out there or soon to be

builders so and they are very impressionable so like you know YC does their impressions so the perfect place to do it yeah YC wants to get the next the next generation of builders to apply so I'm sure it is going to uh it's going to work in uh getting getting more applications in and people are coming

from all over the world when I was in Japan I met uh a couple of people who got into YC Startup School so I'm sure they're there right now just enjoying it and taking it all in yeah and so yeah they even cover airfare if you're So yeah they're they're sparing no expense so 2500 top CS undergrads masters and PhD candidates

100% free every attendee hand is handpicked so obviously you see that Sam Alton was talking today a lot of you know really high-profile speakers they you know they're getting a lot of they're getting the big names in there for sure damn can't believe Elon's willing yeah i I just have this feeling Elon and Sam aren't speaking on the same day no Elon's got to be tomorrow right gotta be

i would love to see them speak at the same time uh all right so yeah that that is uh kind of news item number one so a couple other things that we'll just maybe talk through before we we look at some stuff but the you know it's been since last week since we talked about some AI news so Meta is acquiring a 49% stake in Scale AI so rather than actually just try to buy them they're just buying a a minority of

the company I guess 49% and then you know scale CEO is going to be a meta executive on you know kind of a new super intelligence team or you know they're trying to build a new super intelligence team using some of scale AI to do it so that that's a pretty big acquisition or I guess I don't know if it's technically counted as an investment not an acquisition because

they're just buying uh you know investing in or buying a portion of the company yeah but some most employees will or some employees will get some uh some payout essentially for the exchanges stock yeah which is good which is good we It's an interesting way too because it wasn't like an a normal acquisition so I wonder if more companies are going to try to do this to

skirt around you know some regulatory issues that may come up sometimes when these big companies try to acquire what other companies right yeah like in for at least the funny part here is Meta even if they acquired scale it wouldn't be a monopoly because they suck at AI so they needed this one yeah they needed they they're trying

not to you know yeah i mean I mean and I still like I want Llama to be good i just do right i want there to be good open source models for people to use that are somewhat competitive with you know some of the bigger players so I'm I'm rooting for you know for the open source open source models to win and maybe that's a good

segue into this next one here so Mistrol has announced Magistral so it's the first reasoning model by Mistral AI and so Magestral Small which is like their smaller reasoning model is open source from what I've seen so that's cool another open source i do think that you know from what I've read I don't know if it's that good compared to other

reasoning models but I think it's the idea that it's coming from kind of the EU it is open source so I guess that makes it you know kind of interesting yeah i was talking to a AI coding agent founder and they had a business opportunity like we all do where a big enterprise comes in and says "Hey if you build this uh example or do this feature

we'll pretend to be interested in paying you." and uh but he couldn't use any you couldn't use any closed source model so he was trying to do any number of things llama was not very good um and he did try magestral and it was not very good so unfortunate but there are other models out there that may be better or this one will get better over time yeah did I mean what

about deepseek did they try deepsek they tried deepseek but then it was not allowed for some reasons unfortunate yeah well you know on the one hand I do like seeing more open source models so that's good you know it's it's nice to uh to have and again they do post some benchmarks on on where it sits you know

of course these benchmarks they're going to always pick the ones there there's tons of benchmarks out there right they're going to pick the ones where they do pretty well yeah you can see it's you know in some cases underperforming DeepSeek but in some cases it's close or overperforming so that they're choosing DeepSeek as the

other competitor since that's kind of the bigger you know reasoning model that's open source but yeah always always in favor of seeing more open source models it obviously it's more competition is better for all of us right better models means we can build better things and if if OpenAI puts their open source model that'll only inspire you know now you have something to look at to then improve yours and um

see how it goes so maybe we're on the cusp of a open- source model federation where everyone just starts open sourcing models like Anthropic and everyone else that'd be great yeah all right and yeah so another thing just another news item this one's was it's kind of partially announced last week so some of

this is probably review for some of you you maybe already heard this but 03 Pro is now uh in both pro and team plans of OpenAI and you know there's also like an 80% price cut for 03 so that's cool if you hadn't heard that you know getting cheaper is is a good uh is definitely a good thing uh we do have a comment here from

probably before when we were talking about meta zach's trying to not be bad at AI yes agreed they're trying a lot of things they're trying they are definitely trying um there's 58 of you in here so thanks for joining us yeah so yeah and this is live so if you're watching this right now you're probably watching on X or YouTube or LinkedIn you can just post a comment

we can see them we'll respond to some of them if we you know depending on what we're talking about we'll try to get you on and yeah it's interactive so feel free to Yeah feel free to let us know what you're thinking let us know if we're missing anything what else what's the most exciting thing you're seeing for AI news

and all right so another one this one is you know everyone always uh has probably seen the VO3 all the AI generated videos that have been coming out well there's a new so many now yeah so many there's a new model in town and it is by Bite Dance so this is Tik Tok's parent company it's called Seed Dance so you can see it's a new video

model i have not tried it i just started to read about it a little bit i have tried V3 it's pretty impressive you know you it's all these models sometimes take multiple generations to get the right uh kind of right cuts you if you have some video editing experience you can do some pretty cool things with it but I mean if those are the videos that have been

generated which I don't know if they are those are sick dude compared to VO3 but yeah I mean you I've seen some really good VO3 things though i think again the prompt matters how you prompt it the types of information you give it definitely improves the results right prompt engineering is still a thing you

know you still have to have some otherwise if you don't control you know the direction a little bit then it's you don't know what you're going to get but you can see there's some pretty cool some pretty cool videos in here yeah like that's tight you know but the V3 videos compared to this like if they feel like they all have the same kind of camera lens you

understand what I'm saying like the it looks like it's like the same camera that's like taking all the the videos but this one you can see has a different type of lighting it's interesting i do wonder i do wonder though you know like I just remember I used to do a bunch of uh I guess image generation with stable diffusion and if you you could specify in the prompt like what type of camera

it was taken with and that would actually impact it so I do wonder if you were spent more time with the prompt engineering part if you could get I I have seen some different you know camera styles with with V3 and I imagine like I bet you the prompts are pretty yeah interesting you know i guess this one you can just It tells you the prompt so that's kind of nice oh that's dope so

it's not too complicated of a prompt but you don't know how many times they had to run this generation to get something that was good right they could have ran it maybe that was the first shot or maybe they did a hundred times and picked one that was good and so you can see the prompts get a little bit Yeah longer but still pretty sick that's pretty cool super impressive

like it's only a matter of time because I remember what V3 is like two weeks old three weeks old now maybe a month I don't know I forget now we covered it on the show but there's like this inflection point when those videos get generated and then go on my Instagram stream or my my thread or whatever that

thing is um uh you see all these you know and I I know it's VO3 given the camera and everything i saw some funny ones where like it's like Robin from Batman and Robin he's in like the Batcave he just took a a number two and he's like "Hey don't go in there." Like that type of stuff it's just like very playful but like if it hits your Instagram uh uh feed it seems like it's like now

mainstream you know like because the Giblly stuff from OpenAI like it's just hitting people's non-technical people's streams or in in Instagram and then there people are saying like "Oh you I generated this with Copilot from Microsoft." So like and then there's a bunch of ads now that are being shown hey if you want to turn yourself into

anime use Copilot right so it's like Yeah i guess we're all so early yeah yeah it is uh it is impressive though that the tools that if you are you know it's kind of like if you are creative and you maybe don't have the budget to actually go out and hire a film crew you can actually build some pretty cool things right i think you'll see it in and you've already seen there

was a full commercial you know built with V3 you know it's like it's going to really reduce the production budgets for people so I do think that like these kind of tools are going to be pretty amazing if you're like an indie creator and you want to create some you know some kind of either short film or you

know commercials or Yeah there's a lot of power in your hands now yeah I want to see like I think I'm I'm actually thinking that this uh seed dance is going to be better um would love to see like a music video from like an artist that just is purely AI generated like this yeah you could probably Yeah that's actually a good use case yeah just AI generated for all the Yeah all the parts

of the video but it's you know your music on top of it yeah so yeah there's there's another thing uh seed dance check it out curious if anyone has tried it if not let us know where you uh what you are creating i don't know i don't even know how to try it but I will uh I might do I might try to do some

comparison yeah yeah we find out how we'll we'll do it on stream yeah yeah maybe if we have some time maybe we'll just throw some uh throw some uh generations at it and see if we can get something to actually use it and again maybe maybe it's not fully available yet but we'll figure that out all right and we were supposed to have a guest

they had kind of a emergency uh customer issue that they were working on or customer deployment that came up so we're going to move right into talking a little bit about multi- aent systems because I think this is kind of a pretty good discussion that we can spend a lot of time on yeah couple things from the chat here i

don't know what language that is but what's up um are you guys backed by YC yes we were in YC Winter 25 um and we were the first batch to come back and have demo day so at the Palace of the Fine Arts then Mr frank really like these vibe streams that's a nice way and then he said coding now and it's like AI paparazz

thanks Mr frank for being here so multi- aent shall we yeah so there are two so there probably two articles that we can read through because they have somewhat differing opinions it's like to multi- aent or to not multi- aent it's like to be or not to be that is the question and because of who these authors are Twitter went or went wild just drama as always people taking

stances so um which one should we start with yeah i mean maybe the I think the cognition one came out first right yeah that's the one that started the drama you know yes so and you as it'll be interesting just listen to the titles right the first one don't build multi- aents all right we're eventually going to listen to one that says how we built our multi-agent

research system from anthropic but we'll start with cognition so anything we should highlight around this article yeah so I read this um when it came out and I think first it's okay so Cognition are the the people behind Devon um so you have to take this from the perspective of a coding agent um and a couple like uh just core principles that they uh were talking about is they

believe in what's something called context engineering which makes sense because all we're doing is feeding context into our uh LLMs and so if you're thinking of everything based on context then there's a couple like two or I think there's a couple principles but one is like for an AI coding agent that they need the full

conversation history and um any tools that have been executed before not just individual messages i think that I kind of agree with that you know for sure and then the problem is that if you're using multiple agents in parallel the context gets muddied because a lot of things are happening in parallel but you need the

full picture to actually accomplish tasks so the funny thing though if you scroll down a little bit their diagram here is like a multi- aent that is unreliable in that in what they say unreliable because the task comes in and it does it breaks it into parallel tasks that then have to then join back together back to the agent to do the

next thing and I guess that's kind of setting this up as like not necessarily true because you don't have to do parallel tasks right you could just do them sequentially which is actually what they are suggesting you do so um yeah this is just them saying like the subtasks are being worked on and stuff like that um and so I think yeah

the second iteration is okay well at least the subtasks have the context from the previous task so before it was just they're completely individual breaks down they don't have all the context here they're saying well you could pass in the context so of what actions have been taken so far but then eventually

they end up with this yeah which is still a multi- aent system if you think about it um but maybe the execution is sequential or you know once one uh once one agent is done then the next one begins i don't know i disagree with this whole article but uh yeah and because again they don't really tell us is this the same agent just doing the doing all the subtasks because I think that you could easily make the

argument well in what we've seen is often you get better results from a smaller specialized agent that just does one task well so if you do know that this type of task is needed well you just call that agent and I do get if you can make it you know you don't have to be it doesn't have to be parallel if you can make it sequential and you can say

this step then this step and this step happens of course that's going to be more reliable if you always know the the ordering of the steps well then I think that is uh yeah you should make it in that way and maybe you don't need the parallel you know nature of it and you don't need to worry about combining the results necessarily because you can you

know do it more sequentially yeah but I do think that you could have uh you could still have multi- aent systems that follow the same flow yeah you know when we first built our agent network it worked just like this except there was no context being sent so like 100% the problem exists like where the agents are essentially acting independently in the task execution but what we kind of

discovered and I'm sure the next article also discovered is you need like a working memory set for the task at hand so there's like a memory object that is constantly being written and read from on every level of the task whether that's other agents contribution to that memory or not like it's just a whole

object that just tells you what the state of the execution is and if you have that then you can actually do anything because you any agent in parallel or not can start reading or reading and writing from this memory and uh I'm curious why they didn't you know they don't have any concept of memory here it's just context which is

is memory itself but you know um but there are also right here where let's say you're doing like aundst step execution then you will need to compress memories um you will need to like manage context so like they're totally right in like context engineering being very important yeah and even even with these larger

context models you you there may be not saying there always is there may be benefits for you controlling what gets in that context rather than just allowing the LLM even if you can fit it all in context that might not always be the best case so this is where you you should of course test these theories on these longer you know larger contexts but there may be

benefits for you to you to control how the memories get compressed rather than just allowing the LLM to search for it both for like cost reasons performance reasons quality reasons but you know your mileage is going to vary depending on what kind of tasks you're throwing at it I imagine yeah in cognition's defense uh from my first thing is like most

people don't need wild networks to accomplish a task right so you know if your architecture looks like this that's totally fine you don't necessarily need to have a big distributed workflow or agents or anything like that but some people like I think our natural architecture in our heads is gravitating

towards that because we've been you know we've learned don't repeat yourself single responsibility all this stuff from like our classic training is leading us here but maybe you actually don't need it like monoliths are cool right if if it works for you yeah it's funny so I know we got to go to this other article but I did see a a tweet

and it was just you know it was a clickbaity tweet they wanted to inspire some contention but they basically said "Stop calling agents or stop calling workflows agents." And I think that you know even Sam had a talk on you know agents or workflows why not both so this seems like a workflow you know as well

as an agent right it's like a combination of the two things it's a workflow that calls into an agent in multiple steps i think you end up like weaving these things together quite a bit where you know when you're building an agentic system yes you may use more deterministic workflows and you may call out to an LLM that has some access to tools and you can call that an agent and

you're going to kind of weave these things together kind of like Lego blocks or components that you piece together in your system but but when you can keep the simpler you can keep it the better so the more deterministic you can make it the better so don't use a multi- aent system if you don't have to but there might be use cases where it does make sense yeah we have a comment here from

Adi Singh if the agents are working in parallel is there further room for error in updating the memory object as opposed to sequential tasks 100% right if you're definitely working in parallel and you don't have any write locks then you will definitely overwrite a previous agent's share of the memory the way we designed it at least is like the memory is like per resource within the execution so

it's only responsible for updating its piece but if it's something running in parallel you have to take a lock before you write to memory and then once the lock is free then you can so there's going to be congestion so just because yeah like all this is same it's the same distributed systems problems but in AI

so anything you need to solve those will be applicable here or just do it sequentially if you can yeah and but of of course sequentially leads to maybe slower execution there's all these trade-offs right that you Yeah but I I would say especially with you know these LLMs being not always deterministic right like unreliable you

should probably heir on the side of reliability first like care about quality and yeah more so than performance at the beginning stages and then maybe you figure out performance or improve improve that in other ways because I mean models are getting faster they're getting cheaper like all these things are starting to become more true so I think the quality is still the

thing that people end up struggling with the most all right so let's also then look at this other article so this one's from anthropic how we built our multi- aent research system so in this case they're talking about the benefits of a multi- aent system and again you you got to keep in mind that cognition is building a coding agent and this is a research agent so

they're accomplishing different tasks so you may not you know one size does not fit all depending on what kind of agent or what kind of system you're actually building but some some I would argue though if you're like a coding agent in a codebase half of your job is a research agent right because you're doing all these

open-ended queries on files etc and then there's the writing agent that actually synthesizes than does the response right so obviously I'm biased because I like this article more yeah yeah i uh I would agree but it they do have some you know I believe they have some diagrams here of their architecture for how they so this is how they built into

Claude the advanced research within Claude and so you know Claude chat someone says you know what's on your mind tonight user request comes in or says what are all companies in the United States working on AI agents in 2025 make a list you know you can read the the prompt there so there's the request that comes in so they're using like an orchestrator system or some often call like a

supervisor agent that can then call out to has different tools and can call out to different sub aents so there's a citation sub aent there's a search sub agent looks like maybe multiple you know search sub aents get called in parallel yeah in parallel looks like yeah and then they have you know memory which is

probably you're written to throughout this process and then eventually I'm assuming this lead agent kind of comes back with the final report right it's responsible for putting it all together even Anthropic has agent networks man yeah i mean this is a pretty pretty typical agent network so you know we've

been obviously building out our agent network uh primitive in MRA and you end up with almost always end up with something like this right it's you almost that having the supervisor agent does seem to help as someone that's directly responsible for what the final result right so I guess it's like who do you trust more Devon or Anthropic that's really what it comes down to because like when

this when the Devon article came out all of Twitter was like "Oh yeah everyone like I think I saw some tweet like "Oh any AI infra company is about to lose their funding because it doesn't matter blah blah blah blah right?" But like that's so I mean if you actually read the article they're using multi- aents

too or whatever so I don't know things get polarizing so quickly um in this kind of community that we're in yeah well everyone is looking for the next you know viral moment that captures attention and so if you say something that's a little bit more bold or out there you can capture that attention i think they did a good job doing that

right they did a great job so but they didn't anticipate anthropic coming three days later and being like "Ah screw you guys this is how we do it." Yeah exactly like no actually we do use a multi- aent system and here's the architecture which is almost the exact architecture that cognition says you shouldn't do

think about I mean it's it's a little different theirs is more sequential where theirs is much more you know just like circular right but it it's the same thing this lead agent calls sub aents looks like in parallel it processes the response and sends the final output which is a basically what this is right agent breaks down a task and then combines the result gives a so that says

This is unreliable but yeah again it does maybe depend maybe this is a purely research task where coding agent is not just doing research it's also doing you know writing code and needing to store a different thing so you know again there's some nuance to all of this i I don't even think they're necessarily 100% dis in disagreement on if you actually read the articles but they of

course have differing opinions on certain points yeah and then our users are going to ask us which is better and then we're going to have to be like it depends because it always does and that's not the answer that they want to hear um ever so they're making our jobs harder yeah and and so when you're looking at it from MRO's point of view we we have

three primitives and I it's always hard to tell people when you should choose which one right you have on the most deterministic side workflows then you have agents and then you have agent network or multi- aent systems right and the typical thing we normally say is you know if if you really don't know start in the middle build an agent give it

some tools if you're not getting good enough results you can go more deterministic with workflows or if you need more you know creativity or you need to build more complexity maybe you start to move towards uh in instituting some kind of agent network into the mix but it does vary it varies uh a lot depending on what you're actually trying to accomplish i feel for the user too because having multiple ways to solve a

problem is annoying because you just want to do what's right so you can just get your job done right a lot of people aren't building like AI they're building AI agents but they're not building like whole software stacks for AI they just want to get their job done and the best way to do it you know yeah i mean a lot of people are just building small AI

features you know we say small but you know relatively simple which doesn't mean it's easy but simple AI features into SAS applications and in that case like very often a workflow is sufficient just build a workflow you know honestly like sometimes you don't even need a framework a simple LLM call can do but

once but once you get a little bit more complexity having a system that helps you with debugging and tracing and eventually eval when you get there observability that that can be helpful so as you kind of get more complexity that's where I think just moving towards workflows could be helpful so if you're building something simple like who you

don't need to be fancy just use a workflow you know it works that's what I would recommend too it's because once again we're in this like hype cycle right that you know if if a problem can be solved with AI people want to solve it but if you can actually solve it without it's cheaper you don't have to

pay tokens or anything and uh you just do what you've been doing you know yeah absolutely all right so anything else I want to talk about on multi- aent systems otherwise we have a few more few we can do some more AI news we got a few more AI news items we can chat through just one other thing like Open AI Open AI's agent SDK is uh multi-

aent capable as well so it's not like like this is a real like uh problem space in our community it's just the interesting thing was the the contradiction right so yeah and I mean you know for those of you that have been following this for a while you might have seen you know years ago probably two years ago or whatever when Crew AI

first came out right and the idea was you know and honestly at that time it kind of sucked right because the models weren't very good and so trying to build multi-agent systems with you know they call them crews or whatever like the reliability was pretty bad and so they introduced flows to try to make more

deterministic pieces but now as models get better and better not everything can be a multi- aent system but more and more use cases start to become uh more possible right it's not saying it's always going to be perfect but as each model change comes in you can do a little bit more with these kind of systems that you couldn't do before and so you can kind of unlock more use cases

and so I think we're we're still have a long ways to go but there's a lot more that you can actually pull off today than you could you know a year or two ago all right so there is Let's talk about something else that's just kind of cool that I missed during the news but VS Code maybe VS Code is the new uh the new best editor again oh snap so VS code

seems to be the first editor you know ahead of cursor and windsurf to fully to support the MCP spec so most MCP clients you know cursor and windsurf being two of the popular MCP clients you know cloud desktop is is one as well but most don't support the full spec most really just support tools and so most MCP servers are just a collection of

tools but you know in last week we actually had a workshop where we talked about building an MCP server we talked about prompts which is a part of the spec we talked about resources which is a part of the spec and now and there there's the spec is of course changing it's not set in stone there's more drafts there's

authentication maybe coming to the spec that's in draft status there's more things that I'm not probably not even aware of but as of now it seems like VS Code now fully supports the spec yeah but who gives a you know what I mean if everyone's just using tools and all the MCP servers are just tools it's not like all the MCP servers are

implementing the full spec why does this matter you know that's what I think well I think it matters because once you get like prompts are pretty cool right so you can kind of have prompts that can come in and if your editor now supports autocomplete which is kind of what prompts allow you to do you could

basically start typing and then there's suggested prompts you can almost autocomplete your prompts in your chat interface and if you don't know like users you might not know what's possible through some of these MCP servers so if you can autocomplete prompts that help you along the way then I think it just

gives you that visual indication that okay this is what it can do here's I can click on it and just run this prompt rather than having to type the prompt myself so but the MCP server is never going to implement it if no clients support it so I think you have to have the client support it first eventually

MCP servers will catch up it'll be a while right like this all this stuff moves pretty slowly but if you don't have the clients supporting it why would any MCP server ever implement it that's true that's true and maybe that's the reason why we haven't seen much proliferation in those parts of the spec because most people's clients are like

clawed desktop winds surf cursor etc and then OpenAI is going to do MCP i wonder if they full do the whole spec too yeah we're going to find out yeah is does chat GBT support the full specs does it just support tool calling or tools from MCP and resources is like I still don't fully understand everything about what resources are with the spec but it is cool in that you know before you could

kind of get around it by just having the tool return the context but now the client can just know some of the resource like the context so in the case of um you know like if your MCP has access to information maybe it's user specific information you can just return that as resources so the the MCP client just

knows about whatever user data you need rather than having to call a tool to get the data so it's just a like I don't know how well it will work yet like I don't I think there's some risk with clients implementing the specs like or reading the specs slightly differently like how do they use those resources every agent's going to be different

but it is uh I do think it allows you to kind of like pass more information into the code agent a little easier yeah and also MCP is so new like what like seven months old now eight months or something and then in the prime spotlight for like four months like right now so yeah i do yeah go ahead sorry one last

cool thing if you're using Windsor for cursor it's based on VS Code so it's not like you can go back and like it's going to be a completely different experience it might be though for a lot of things but maybe for this it'll be super chill yeah and I imagine Cursor and Windsor have got to be right behind right yeah they're not going to let VS Code like

own the limelight too long like they they uh they have a lot of users paying them a lot of money so they're going to eventually add support some of these MCP servers will get updated but I am curious if you are watching this how many people are actually using MCP servers in their editor because I bet

it's still a pretty small percentage like we're in this hype cycle echo chamber where MCP is everything but how many are you actually using obby if you go into Windsurf how many MCP servers i just have one which is the master doc server i Yeah I have uh three monster docs and then two others that I was that I was building for fun so that's it like

I I think that it is you know again there's a lot of hype around it i think the hype is warranted because I do think you can do a lot of cool things with it but I don't know if like a lot of people outside of developers are using MCP yet maybe some people are using with cloud desktop maybe now with chat GBT

supporting it MCPs start to become even more interesting because chatgbt is obviously a big consumer application so we will see but in our actual agent applications we do use MCP for things so it's not like we don't you know like Slack or like all these things off the shelf that are you don't want to write yourself um that's

where I see like the most value right now yeah yeah if you're building an agent yeah maybe less so in just the clients that you're using every day but yeah if you're building an agent it makes does make things a lot easier oh we got two two people who told us what's in their MCP except I need to There it is yeah I think you're we were both turning it on at the same time paul Bunker thank you

for that comment mra and one for accessing files on Machine that's dope and then this dude armor only context 7 for getting docs which we've heard about that uh that before so yeah yeah context 7 for those that don't know is a pretty cool way to get access to a whole bunch of documentation de developer

documentation really easily so you can turn on contact 7 you can get access to tailwind or someone has added the monster docs to it i don't know who but you there's a whole bunch of different developer tools where it has access to that those docs it basically makes your editor a little bit more of an expert on

all these different tools because of course these LLMs don't have the most up-to-date data or content for all the these documentation so it is helpful uh I've I've seen you know I do think that the last time I used it sometimes it struggled with searching and finding stuff i mean but even RMCB doc server

isn't perfect so I think you know everyone's searching is not an easy problem right search is still a complicated problem and how you actually search that depends a little bit on how you you know what the prompt is all all that stuff but it is a pretty cool tool all right so we have a few minutes till our till we go into security corner obby tell the

tell the people the people want to know so you're doing some traveling let's spend a few minutes let's talk about all the things that you know that we did or I say we you did i just watched from the sidelines and and some of you all maybe did as well but yeah what what were you doing what and who are you with nice so

um the trip started Oh dang almost like a month ago um Tony from MRA me um and me and Ushwin who's also on our team we flew to Japan for two special events the first event was a like a meetup at the Layer X is their company called Layer X they had a big meetup over 200 people were there and I have a bunch of stories along the way um so I'll try to be quick

um then we went to a CTO gathering of essentially Japan's top companies everyone there had revenues in the hundred millions like in like they're big and they're like I met these two guys they were like asking me all these questions and I was like "Dude how big your company?" They're like "It's about like 500 people." And I was like "Why are you asking me questions dude our company is like way less than that." Um

but was a really cool um experience we also met with the NRI which is Namura Research Institute they're one of the biggest companies of and billion-dollar companies in Japan and they really wanted to know like how they could you know use MSRA for client projects to do because they do a lot of consulting they're a big company like that

consulting is a big deal in Japan as I we've learned you know um so we met with all those people i dipped and everyone went back home except for me i flew to uh Belgium and I took a train to my our another Amastra's house ward Ward Peters and we co-worked for a couple days um if you guys saw the ADA video and us look working on the the horses and goats

in the backyard uh that happened we did a bunch of streams as well while we were hanging in Belgium then my little brother Nick came and visit or came to meet us in Belgium so we met up we went to a small village next to Ward's place because no offense Ward but your town is small and boring and we were there's like a city right next to it that is small but not boring so uh we did that

and Ward came over because it's only a 20-minute car ride for him then we hung out again co-worked some more um also went to France to see Marvin from our team he lives in Strawburg y'all if you were watching the EU streams we did streams there as well and then all of this kind of culminated in uh a wedding that we had to go to uh for Tony who's a part of our team as well uh

so first there was two weddings so first wedding was in Greece so we flew there wasn't any work happening there i did record a walking video in Greece and I'll release it later the the topic is like I'm gonna show you something that AI can't do which is get married so like I have a whole video for the wedding and

stuff and then we flew to Helsinki Finland which that place is wild because the sun doesn't go down till like 12:30 a.m and it comes up at 3:00 a.m so me and my brother we were screwed our sleep schedules were terrible just couldn't get any sleep because there's just sunlight all the time um there was another wedding there which I officiated

i hope I did a good job and then came back home so that was the whole trip very MRA focused though met meeting all the teammates it was definitely a lot of fun can't say it wasn't um so yeah but I imagine you are back or glad to be back in San Francisco dude yeah at the end of it I was just so ready to go home

because I hadn't slept in my own bed for a while and I kind of you know I missed my co-founders a lot i I know it's kind of cheesy to say that but I truly did because uh we usually see each other way often than you know I was gone for like almost five weeks so we each month yeah and time zones are tough too and yeah

there's not not a lot of overlap but I will be on your doorstep in about I don't know 20 hours so I will be there soon also for the viewers Shane and I have this little tradition that we're not going to be able to do tomorrow where I pick him up from the airport and we there's a probability if I'm going to give him a hug or not something like that but I don't have my car because I flew directly to San Francisco and I actually

left from LA so we won't get to do the tradition tomorrow unfortunately yeah it's fine i I'll forgive you this time but it has been like six weeks so you know you know we we will we will see each other soon and yeah this week's going to be fun you know we have so I'll be speaking at uh an event an AI meetup

at the elastic office where I'll be talking about building a personal assistant agent with MRA so that's uh tomorrow and then on Wednesday I'll be doing a short demo at AI Tinkerers which is also in San Francisco talking about how we built the MCP course for Maestra so it's going to be a busy week we're

gonna have a lot of fun yep and we're planning the future of MRA so as we always do some more some more road mapping but we do have a special guest you know it's been a little while since we've done a security corner let's bring on Ally ally what's up nothing much what's going on just just thriving and surviving yeah can't can relate for sure but

excited to do another security corner yeah so today you know and I'm going to be honest I didn't read this full article but I did have uh GPT40 summarize it for me so and then I read the summary so we are going to talk about you know it's actually a paper and it was posted on Simon Wilson's blog so I should just share my screen we'll talk a little bit about this and we'll just you

know spend a few minutes chatting but it's called design patterns for securing LLM agents against prompt injections so it talks a little bit about the scope of the problem as long as agents and their defenses rely on a current class of language models it's unlikely that general purpose agents can provide meaningful and reliable safety guarantees that's pretty bold problem

you know I think a lot of people would probably if you really think about it it makes sense though right so what kinds of agents can we build today that produce useful work while offering resistance to prompt injection attacks so they talk about different patterns and so there's I think six different patterns that they they talk through

here and yeah I'm going to try to just pull up the quick summary so we can talk through it any first highle uh thoughts before we before I give the high high level of all the six different patterns yeah for sure um I had some high level thoughts reading this and just like you I read this like not long before like doing this so I didn't fully read the

full paper itself that they linked to but I did read most of this article um my like one largest takeaway I think from it was that the risk is like never zero um like you can do everything you can you know try to have an LM guard in place or an AI security runtime product or implement one of these design patterns but not any one of those methods in and of themselves is fully

foolproof which reminds me a lot of cloud security and how you know that's sort of a shared responsibility model and um you know simply turning on like precautions with your cloud provider that's not a foolproof um you know model of app security for your application so it's much larger than that and I think that's sort of the theme here with AI

agents and that's my biggest takeaway yeah all right so let's talk really quickly on the six different patterns i don't know that I'm even going to fully understand them by just reading them but we will do our best to decipher them as as best we can so the first one is called the agent uh the agent selector pattern so

essentially the agent can call tools but doesn't see their results so it's more like an LLM driven switch that prevents the tool output from like changing the LLM's decision and this one is interesting to me because if you I know I've I talked about it last week i talked about a lot but we built this course using an MCP server and the whole

idea around what the course does today it's just a bunch of tools but it tries to get the code agent to think it's an instructor so we are literally using the tool output to kind of change the agent in this case like if it's cursor or winds surf to to treat it more like it's instructing you rather than doing its

normal job which is to write code for you so we are literally in some ways kind of like prompt injecting back to the editor in order to actually get it to treat it you know treat the itself like an like an instructor so it's trying to teach you the content and it can help you write the code of course but it changes the mindset of the the

agent a little bit so we are essentially using prompt injection to do this so if if you didn't have access to the tool results of course you couldn't change the way that the agent works or if they guard rail us and say you can't be an you can't tell your agents you're an instructor then we're screwed yeah it's

why we call the course experimental because it's definitely not the intended use but it does work relatively well depending on the model that you select uh you know and certain you know cursor works pretty well windsurf is a little bit more verbose but it still kind of works but but yeah we're basically trying to

like in some ways like hack the agent to become an instructor for you instead of just being a you know a coding assistant all right so the next pattern is plan then execute so essentially the agent first plans a set of tool actions without seeing the tool output and then it executes so it separates the plan making from the execution so it at minimum malicious

instructions in the output can't influence the behavior this one sounded like an LLM workflow to me i don't know if that sound what it sounds like to you as well yeah very much so yeah well it's almost like the LLM like the first step is plan make the plan and then it's like okay so then after you make the plan you can execute the tool calls but instead of

just allowing you to you know making the plan as you go and and feeding the results back in all right so then third one is the LLM map reduce pattern so basically apply the LLM over chunks of untrusted text independently then aggregate safe summaries so it's useful for large input and helps prevent crosscontamination

this one made sense to me up until the aggregation parts like what decides if it's safe or not like where's the LLM as a judge I guess in that workflow to decide that you know the one agent that did go over one malicious file that that was malicious or not you're going to need a judge and it does say involve sub agents so maybe the

judge is part of that you know or maybe there's like sub agents that help judge files that were judged relevant are then aggregated and sent so there there must be some kind of sub agent that does some yeah sub agent that responds with the boolean indicating whether the file is relevant or not so yeah it's it's like a distributed

pattern for you know like analyzing large chunks of text and determining if it's safe but yeah there's it seems like it can be a little complexity to to implement but there's a lot cuz like you have to like if you're doing a stream you have to you can't just stream all the text right you have to collect the text in chunks see if it's okay then

allow the stream to happen it's kind of annoying actually but uh it's secure I guess yeah does that impact latency oh for sure yeah because a lot of stream text is like one word athe the from from the from the where you know you got to wait till that so you're adding I think all of this adds latency every single one probably so but that's okay you know to

be to be secure so I mean it is and isn't like it's got to be a usable product for sure like I think when I was reviewing a lot of like AI security runtime products back for a paper I wrote in January February timeline a lot of the things that we looked into was like latency and if it you know introduce a bunch of latency to the

products then if it did um then you know they didn't want to use it there's an article out there by Dropbox that they were evaluating a product like that um and they listed out their top three requirements as an engineering team um and latency was one of those three yeah i think it I think it varies there are there certainly some applications that can be more latency sensitive right

you know if you kick off a deep research task in Chad GBT for instance you know it's going to take a while so if it added a little extra latency in the grand scheme of things it's probably not noticeable but if it's a autocomplete for a code editor you can't add extra latency there right you have to you need the autocomplete to be snappy and fast

and so I think it it definitely there there's a spectrum for sure that's a great point i think it's true like for voice agents also like it's super weird to be on a phone call and then have like a bunch of like lag and then it's like okay maybe this isn't a real person or this is kind of strange yeah yeah i mean voice agents are probably the most latency sensitive

application like but I doubt they're doing prompt injection patterns at all yeah i think Yeah I'll put a bet on it i think you could probably I bet if you tried you could probably prompt inject any voice agent that's out there because they you know they're so latency sensitive i mean we have people that are building voice

agents or you know voice agent type capabilities in MRA and they're comparing like detail latency metrics for every model right because the the results of the the actual content the model produces is important but maybe more important is like even if it's one is 80% as good but twice as fast like you would take that because the latency is so important and so I imagine that a lot of them don't have they don't have

the extra latency time to be able to try add security right now yeah I totally agree and I think that was like one takeaway of mine also from this article i think at the top they said there's no magic um fix for prompt injection as is um even though I know a lot of like runtime security vendors kind of claim that they are that um I'm not sure if if

you know AI runtime security is the solution or you know these these workflows or these patterns are or probably a combination of the two but you know one of the most notable voice agents um the Vader NPC in Fortnite that started off really bad um with like a lot of like inappropriate output and stuff and falling victim with a bunch of

different prompt injections and then it suddenly like it was able to be fixed or it wasn't doing that anymore so like clearly there is some sort of solution out there like it's possible totally yeah yeah it's I security I think is always a layered approach right like there's always multiple multiple steps it's not necessarily just one thing all

right so the next the fourth pattern the dual LLM pattern so this one says like a quarantined LLM processes untrusted input it sanitized output becomes an opaque reference for a privileged LLM and this shield prevents malicious tokens from ever reaching sensitive logic i don't know what that necessarily means the quarantined so

they like so this is a dual LLM approach is how most guardrail systems work where like the guardrail itself is an agent or an LLM that is a smaller model that can judge the output so essentially the request goes to both LLMs and they both start executing so you still need a good execution environment right you can't like not have that but like the guardrail LLM or this quarantined LLM is

going to like essentially judge if it can actually continue the operation or not right and if it can then you're you didn't lose any latency because you already started the request with the LLM that's going to serve the response right and then if you do find some you can trip wire or you can just cancel

the request so it adds complexity for sure but it's probably the right thing to do yeah and I think this is what a lot of the low latency guardrails applications are doing you know whether they block it and wait on the result or whether they run it in parallel and then they they wait for both to complete before they pass in the result but

they're basically doing some version of this right yeah honestly none of these are like this one's a great solution i I like this one a lot streaming is what everything up right like if you didn't have streaming then no one would give a about all of this uh these like different patterns i mean they wouldn't care that it causes latency or whatever but because it's hard to fix a

boat while it's freaking going like exact exactly what a like a stream is then you know it just gets harder to do yeah exactly streaming uh streaming is always a pain every time it always comes back to streaming all right the next the next the fifth pattern the code then execute pattern so apparently this is discussed in the deep minds camel paper but the

LLM emits structured executable code so this could be you know Python or TypeScript code or whatever instead of direct text so then the code is validated before execution so it's reducing the chance of like hidden injections and then the code's executed right so it's actually saying that no you have to write code as your that that's your response and then you you validate the code and then run the code

okay this I guess that one makes sense to me um totally that one kind of reminded me of like run Python the MCP server that Pante created with the Piodide yeah it's it's funny i think all of enterprise agents will have to run their LLM requests in a sandbox environment and then you're going to have to have like a proxy in front of it for actual

user interaction i mean I think I when we we're at that OAS thing Ally I think I thought I saw architecture diagram kind of like that for like a sandbox execution um that's probably where we're all headed so that sucks because it's such a pain in the ass to do but it's important so yeah we'll see if patterns improve there

because yeah it sounds like kind of like a bastion which like that's sort of an antiquated model and we've got better improvements from that since then all right the sixth context minimization pattern so this one is basically supply as little context as necessary to the LLM to minimize ex because minimizing exposure shrinks the attack surface and

limits potential instruction leakage so even Simon Willis says "I'm slightly confused by this one." So all right so if a user's prompt is converted into an SQL query which returns raw data and that data is returned in a way that cannot possibly include any of the text from the original prompt the chance of prompt injection could be eliminated so

again I guess it's you know in some ways making sure that the context that gets passed into the LLM is minimized so it's getting as as little context as possible that actually gets passed through yeah if you have like poisoned context too then you would need to have some one you have to get rid of it right but

you'd also need to have like some safeguards on the inputs to your request like your whole memory chain would need to have a filter on it like throw out PII throw out all the stuff that shouldn't be going in there yeah I know a lot of like uh Chat GBT rappers that claim to remove like PII and then throw the I'm HIPPA compliant sticker on it because like OpenAI isn't like HIPPA compliant um to my knowledge

well it's getting saved somewhere exactly well it's getting saved in a HIPPA compliant database so okay then everything's chilled at that point yeah in theory and then they so this paper then introduces 10 different case studies which I'm actually pretty interested in this i'll probably read through some of these because it's talking about how to use these some of these design patterns

but they're also like practical examples of building agents right like an operating system assistant I guess SQL agent email calendar assistant customer service chatbot booking assistant product recommener I mean a whole bunch of actual practical agents these are things that MRA users are trying to build right i'm sure you know if you're trying to build an agent it might fall into some kind of category that's close to these things i think I

think uh it's nice to read case studies on how some of these some decisions making some decision-m happens when building these types of agents so uh definitely if you're building something similar to one of these it might be useful to read the paper and you know unfortunately they just give you the Python code you know instead of the

TypeScript code but it is useful to see just actual practical code examples of how some of these things are getting built all right and there's some other stuff you know closing thoughts and you know actually a link to the actual paper to read but definitely some interesting thoughts around prompt injections and how you might be able to use some of these patterns to help at least make

your agent a little bit more secure i don't think there's a you know there's no silver bullet as we've discussed but some of these things could help you know coupled with maybe some of the you know security providers that are out there that help in different ways you know as we mentioned kind of a layered approach probably there's not one thing that is going to save you but a collection of

these things might make it at least a little bit more secure for sure yeah we call that defense in depth yeah I did look at the white paper and I thought one of those use cases was like a medical diagnosis chatbot which I thought it was interesting and they use like context minimization which makes a

lot of sense because whatever the user says like they don't keep that in the the memory of the interaction what they do is they summarize the symptoms get an actual diagnosis from their knowledge base they do a rag on a bunch of medical uh maybe medical journals or I don't know textbooks or some and then they return that result uh to the doctor

right because it's the doctor that's using the bot to take all that stuff so the doctor actually doesn't know what the user um even said or anything they just know what the based on the symptoms what the diagnosis was so it's pretty interesting i would read the white paper for anyone who's interested in reading

white papers yeah so you can see this one i'm assuming you're looking at this one medical diagnosis via an LLM intermediary yeah um but yeah they definitely encourage you to you know for those of you that do like reading white papers or you know my my preference is I will often take these white papers and just uh throw them into notebook LM and have them generate a podcast for me and then

I will listen to that 10 to 15 minute podcast around uh the white paper and learn a lot without actually spending the time to read at all depending on you know if I'm hand need to be handsree or if I'm traveling or something sometimes that's one way one hint that I I like to do that helps me consume some of this information without actually reading all of it because there's a lot of white papers coming out it's hard to keep keep

up with all the stuff but that's why you tune in to this so maybe you can keep up in less time all right and yeah we posted the PDF so there's the actual PDF there if you do want to get the white paper and read it yourself or drop it into Notebook LM or wherever wherever you put your PDFs and with that anything else going on

Ally that you want to talk about today um no nothing really top of mind all right well we appreciate you coming on and chatting with us about some security things and we'll we'll see you again here in the near future I imagine awesome thanks thanks Obby see you see you see you and then it was Yeah and then it was two

just you and me again that was my first security corner yeah we we do all kinds of stuff sometimes we'll we'll show code we'll talk about Yeah various various security related aspects security is important especially as you get to actually productionizing an agent right it becomes more and more important especially the larger of an organization you are the more security you need right just to feel confident in

what you're actually deploying and it's hard to feel confident in some of these AI systems right now so I think it is important to talk about it consistently and to you know slowly figure out ways and tactics to improve the security of whatever we're building yeah and it's very frustrating too to the user because it depends is the answer to this one too like how do I

secure my agent well it depends you know and not just one thing all right and with that so we're going to talk through just some some stuff that we are working on at Mastra we do have then um MRA user that built something pretty cool that we're going to have on here uh hopefully pretty soon maybe in 15 or so minutes but we did want to talk about

something that I think Ward was working on where we we maybe wanted to show some code and show how it is actually possible we had a question too it came up on Twitter it's come up from you know a customer that's trying to use Maestra where they basically said "We're using Superbase we're kind of all in on

Superbase but we can't run Maestra on Superbase Edge functions." So we thought about it and we're like well why not we should be able to you should be able to run Ma at least you know some version of it in a Superbase edge function and so we have a working example that does work abby you want me You want me to pull up the code i got it all right i got it can you And let's also Yes we can see it

hold on one second let me stop sharing i'll share the link with everyone here is it public oh I forgot to check that i don't think it is yet uh it's public it is it is okay yeah cool so let me share this okay okay so we were talking to um a couple customers who made Superbase edge support and we you should want to do this because I

mean we should want to do this that's why we're doing it um but Superbase Edge Functions runs on a platform called Dino and Dino has two versions dino version one and Dino version two duh the thing is I got to tell a story before we get into Dino because we were at a company that spent a lot of time trying to get

Dena working on their platform and it took like many many many many months to do so um and so when we entered the company the like the remember Shane like the nightmare that people were talking about they're like "Oh after Dino we're not taking chances on third party software ever again." Which screwed up our job

and it pissed me off right um so I don't have a like a very pleasant like memory of Dino in my life what about you yeah the the same yeah I you know Dino had I would just say some reliability issues and it seems like you know they've gotten better from what I've seen here but you know this is a point this is a few years ago a few years ago there was

some issues and it was incredibly frustrating so yeah I have a negative perception of Dino because of my past as you probably do as well Abby but you know the people want what the people want and people want what they want and we will give them what they want and they want Dino support because that's what Superbase Edge functions are running so we will figure out you know

how to make it work with Dino yeah so to get into what we did so first of all we had to go through all of MRA so MSRA itself was built for node um and designed to run both in a serverless node environment or longunning processes which you know we believe that in AI world you should probably use longunning processes if you can um just to reduce the friction that

you may have when starting out so superbase edge functions are on dino one I believe so what we did here is created an example we're not necessarily going to make this like first first party support yet until we have more users with that they expressing that they really need this because then we'll just create a dino deployer right so you can add that to your MRA instance it'll do the

right thing when it builds and deploys and all all that will be good what we wanted to show in this project is that you could essentially build a monster app deploy it to a superbase edge function and then in a UI use our to like visualize workflows and it was cool because in to to make something dino compatible you have to have a bundle ruski on your team and if you all

don't know what bundle ruski means is someone who knows how to bundle JavaScript very well because we have to support all types types of runtimes not just freaking Dino so we had to take our node runtime and then do a bunch of trans transformation at the build time uh if you all don't know Dino the imports are different if you want to use a module from node you have to use like node colon it doesn't support port

process m man there's a whole bunch of that it doesn't support so we had to go and make those changes which is fine we we did it so in this case you're running the application within superbase so I'll show this is like a normal superbase folder that you can get this just has all you know if you like superbase in it you can get this and in the functions directory so we have this

dino json you can see these are like the these are like the mappings so if you want use like npm module you have to do mpm colon etc and right now we have like our uh like this is like a snapshot version of what we're building so don't really pay attention to that other than it's just a specific version of MRA or cooking to use here and if you want to

use we use pathbased imports which is another thing you have to kind of like figure out uh you can't just do this i guess you could but like we need to we had to like map like what our path imports are to the Dino ones so as I said like mpm colon package name is kind of like the syntax there oh what happened here where' the CSS go on get up there it is

um so then we you know MRO uses Hono to as its server you can execute Hono anywhere especially even in Dino so we had to set up a and you know if you want to do this yourself you'll need to for now you'll need to set up your hono server and then you can add um one good thing we did in MRA server we have a package called mashra server what it

allows you to do is if you don't want to use hono or you have to use hono in a different context like this you can import all of our base mo server handlers these are not server specific but they do all the server functionality that is in our MRA server so here as you can see you can set up your own honer in a superbase edge function and then you can just call our handler and pass in

and then it'll give you the same functionality that that same API handler is doing in the normal MRA server so if you want to build express server KA whatever you can do so without you know having to just use MRA the whole server so that's what we did so this from this perspective this was the easiest part right just setting this hono stuff in superbase the hard part is actually in

the core framework and then you know we did all that so I'll just show you this is as you can see a typical MRA project as you can see process has to go from node process you have to inject some stuff the workflows it's a test workflow all of the same stuff but uh we made it denote compatible internally in MRA core

um and we will release it um soon should I look is that that's pretty much it to to show if you want to play around with it you could just use this as a like or copy this or look at how it's done and then try it yourself and I think this is just a UI that we showed that you know you can get all of MRA's if you like MRA UI and stuff we

export the playground UI but we don't necessarily tell you how to build together and that's fine because that's not what we're trying to do right now but if you're adventurous like we are in this uh example you can import monster playground components and then put them together just like they are in our playground we get a lot of requests to

do that as well so this kind of shows like two ways of doing that other than that happy hacking if you're really interested in this yeah and this is a pretty typical approach where we we'll make an example and show show you how you can do it and if enough people ask for better support we'll we'll add more

right we'll we'll improve it and make it a first class deployer at some point but we like to show you what we're what we're hacking away at what the team is is working on because we you know this honestly came from one one of our users one of our customers that wanted wanted this and so we're like let's build it and we heard another a second person raised their hand on uh X around the

same time that said they wanted it so we'll build an example and then if enough people like it maybe we'll we'll add more support and make it even easier uh so there is a question and I do you guys know of any good examples of STS i'm assuming is that like speech to speech maybe like real-time speech I would guess and then front end to

separate the master back end yeah um yeah uh we don't we don't have any good examples uh yet um that's because if you really want to do good STS you're going to need to use WebRTC on your front end to then go to Monster's back end if you're using a Monster back end you can also use um you can also use uh sorry I was laughing because awards got

sorry Joseph and has great the developer who created this one of them great hair yeah great hair back to Joseph's uh thing oops sorry go ahead Shane um yeah we need to still work on this WebRTC to Masha server connection openai real time does have WebRTC Web RTC support and I would recommend doing WebRTC um if you're like depends on what

your front front end is but if you have WebRTC support then that's where you should be doing if you don't you can do STS where you have websockets and then you know you're you're in that game that's probably easier to pull off right now in your own server if you're trying to use MRA as the back end we took a little break on voice because

there's a thousand other things to do but we will return to the voice game one of these days but not too too long from now so also Joseph thanks for being here again i think I remember you from a different stream like last week or something or the previous week so welcome back yeah so with that so we we were originally going to have a guest come on and demo

uh he did send me a message and say his demo was broken so we do uh we will have him on at another point we do have a here's another comment oh yeah we have a we have a bug in the latest that's actually in alpha now um something was wrong with our stream encoding on the HTTP side uh we pushed a fix for that this morning it will go out into stable release tomorrow um so yeah

sorry about that the it was introduced in 010.2 we don't really know how or what changed yet that's where the team is kind of investigating right now timely question though because yeah it should be fixed tomorrow the streaming for some reason is just collecting all the tool calls and everything and then responding at the end so not really streaming you have to change your chunk type or your chunk

encoding to chunk or character encoding to chunk one of those encodings but um if you have patience uh it'll be fixed tomorrow then Ysef said uh thanks yep thank you yeah no thanks and hopefully we can get some some better support there too all right dude dude's demo broke that's what happened sorry we were Is that what happened yeah so yeah hopefully

hopefully the demo wasn't you know trying to stream something answer the phone it was though i I doubt it um but you should have answered that phone live on stream and said "Hey yo I'm on I'm on a stream right now." Dang um from Nina Patit could you cheese please check GitHub issue of rag i can't create rag with the last MRA version uh yeah

we're also aware of this um should be uh we'll fix it in the in the release same thing I told the other dude uh we're on it yeah tomorrow more improvements more improvements uh fixing a few regressions moving moving the ball forward but yeah tomorrow Tuesday is always release day for those of you that that do not know

we we try to release on Tuesdays every week so there's always new things coming out and we then highlight it later in the week when we do the change log oh Discord issue yeah uh with that yeah anything else you want to chat about today otherwise we could call the stream early you know we've been going an hour and a half 90 minutes is that's still a solid stream yeah and 162

people in there hope everyone had a good time um but I think as the chat has told us I think it's time for us to go fix some bugs yeah always more to do thank you all for tuning in uh we've been we've kind of been changing our streaming schedule up a little bit so we're we're still going to do these AI agents hour live streams every week

typically on Mondays around noon Pacific time the rest of the team is still going to be popping in doing more actual live coding streams so we decided to split up the the guests and the news into one stream and then the team will come on occasionally and do like coding live streams as well so I think you know Tyler and Daniel are going to come on on

Friday and try to do on Fridays and do a live coding stream so if you want to see live coding we'll have some of those streams for you but you'll know kind of going into it what you're watching you know and then if you want more news and guests and kind of like what's going on in the AI world the the AI agents industry I guess if if you can call it

an industry that's popping up that's what the stream's for yeah oh we also forgot to say something um one of our guys Marvin who you all know from stream just had a baby so if everyone can give him a congrats and well wishes and whatever positive energy always helps because he's not going to be sleeping anytime soon so yeah congrats Marvin get some sleep good luck

and um Ward when's Spotify we're already on Spotify we're a little out of date but we are on Spotify right now so if you want to watch it on Spotify it's a t it's a little delay because there's a manual process and the manual process is me being the bottleneck from stream to upload to Spotify so we're a little behind uh but old episodes will

be slowly getting fed into Spotify so if you'd rather watch this on Spotify and not watch it live you can do so and we'll do EU streams as well we just have to like well one dude has a baby and one guy has three goats so we have to figure out you know the right times and stuff but yeah we'll come back yeah and those will

likely be you know kind of the the live coding flavor right so all you know all let us know you know definitely let us know if you like kind of splitting up these formats do you prefer seeing the news and having guests on talking about what they're building in AI and seeing demos do you like the coding live

streams but hopefully the the splitting them up you'll be able to know when you're jumping in are you are you in for some live coding or are you gonna in for some more news and guests and updates agreed Amar congrats and yes please Ward yes please subscribe to Spotify and why aren't you just on the stream right now with us what are you doing what are you doing right now

ward is feeding his goats that's my guess feeding his goats all right well uh thank you all for tuning in we'll be back again next week we'll have some more live streams we always uh one other thing we live stream on Thursdays we do a workshop each week so we try to live stream that uh so that's a little bit more there's a mix of of coding in

there and some education so we we do a live stream our workshops as well so hello I'm Jed thanks for joining us we are just about to wrap up but tune in next week you can always check out all of our stuff on all of our past streams on YouTube so go check out the master YouTube it's right there mostra-ai on YouTube if you're not already please

follow Abby on X it's Abby please follow me as well SM Thomas 3 thanks for tuning in to AI Agents Hour we'll see you next week see you peace