Cavemen With Fire: Governing the AI Agent Fleet ft. Logan Kelly

Everyone's shipping agents. Almost nobody is asking what happens when a hundred of them go off the rails at 100% utilization while a customer is on the other end. Logan Kelly — CEO of Waxell and a guy who's been punched in the face by real-world sales operations — joins us to argue that a dashboard isn't governance, it's an autopsy, and that most teams are "cavemen playing with fire, hoping there's a bucket of water somewhere." We get into why governance actually *accelerates* agent development (brakes make F1 cars faster, not slower), why the real competitive fight isn't OpenAI vs. Anthropic but governing your entire agent fleet across providers, and why vibe-coded production apps are the next wave of preventable disasters. If you're building, deploying, or betting your company on agents, this is the conversation you can't afford to skip.

Transcript

0:08

hello everyone and welcome back to another episode of the AI Rebels Podcast as always I am your co host Spencer and I'm your co host Jacob and we're very excited to have Logan Kelly here with us who is a man of many talents uh he is the CEO of Waxell AI and co founder of Call Sign both AI based platforms and tools one of the reasons I've been excited to talk to you Logan is I feel like everyone is building agents right now and if they're not they should be in my opinion but very few people are actually asking the hard questions of like what's next like how do we monitor these how do we once these become common and everyone is working with agents what do we do like how do we where does fault lie if an agent goes awry and everything between so for our listeners Logan has a unique background where Waxell is a governance observability platform that kind of sits above your agents and make sure they're you know what they're doing and then on the flip side call sign is an AI powered sales platform is that a fair way to yup describe it Logan OK yup um involving agents too so it's like you've got both things going here which I'm sure is a nice blend so all that being said thank you for coming on Logan we're super excited to have you thanks for having me guys can you um I feel like everyone we talk to has a different path to AI and how they got where they're at now and it's fun because it's like in the last few years it's been so fast can you walk us through how you got to where you are now yeah so I it started I I actually started in the car business in my my 20s I had a couple of stores and it really started with like the idea of like being able to use you know at this point like deterministic automation and all of this to to sell more stuff right and and that was super exciting you could build processes make sales people more efficient and then you know kind of got out of that and spent some time on uh building out AI before like LLMs were common place like they are now in the nutrition sports nutrition uh side of things while also uh doing a lot of work on the B to B you know outreach side you know and and that and uh and so it was really kind of like all the the kind of common thread has been can we use technology to make you know real life stuff more effective and and get those outcomes uh cause like growing up as a a sales guy to start it was like you could do you know nobody cares about the effort you put in right it's all about the outcome right and and so that was a lot of a lot of like what I've been trying to do in all of the things that I've done is literally take technology and make things more impactful process wise and and that kind of thing I like that I'm curious this is slightly aside but the name Waxell AI where did the name Waxell come from yeah I'm actually from Alaska uh and I've named a few companies and and I was like as we were building this I was like this is a huge undertaking and uh Waxell is a a ridge in a huge like glacier field and it was also oh awesome the guy bearing like if you heard of like bearing straight bearing C this kind of stuff his like second in command who when Bering died when they when they had the expedition to Alaska from Russia he was the dude who got uh the whole crew uh back so uh that was some every Alaskan has like some crazy like super nichey specific knowledge about Alaska and that that's like one of my favorite stories he's like the Russian Shackleton you know um wow so and he was Waxell was his name yeah his his last name was Waxell yeah probably not pronounced like we pronounce it in the US uh right but yeah it's just a cool like niche story that's like super badass that you know I figured and it's very yeah uh pertinent I mean he was like the fallback you know yeah Barry and died and he was like he's the one that caught it he's yeah he saved it yeah yeah yeah he was he was cool yeah that's right that's awesome so I had to ask with a name like that I knew there had to be a cool story behind it for sure so I mean I'm curious with Waxell as we're on it you mentioned that you you know selling cars lead gen like these are not the typical paths we have seen well I guess we have seen it which is cool with AI but it's not as common often times it's someone who has a background with software engineering something like that yeah whereas you're kind of coming from a different route has that how has that been a competitive advantage for you yeah it's interesting like when you live I agents are so you know they're you you have this thing that you build and then you know like you guys are building agents too like you put that into production like or or you put that in on real data like you're not there's nothing abstract about what it's doing right there's no there's no like oh you know uh some server log or something like it is impact it's either delivering information to a customer or it's delivering something internally and so the the kind of the outcome focus that I've always like had in my DNA is it's like a gentic governance becomes something that it's the only way to guarantee one that things aren't gonna go wrong on the other side that you're actually gonna get the the outcome so I feel like the people who are agents are this interesting place where the people who are practitioners right like you know doing the things on a daily basis I think are are uniquely positioned to build this stuff and and also you know one of the the core tenants of of Waxell is we want governance and the ability to understand and and um control what agents are doing not to just be an engineering function but to be a function that the entire organization has an easy pane of glass that they can that they can then view this so coming from that operational like day to day like you're getting punched by a customer who doesn't like what happened right right or you know as a startup founder right like I'm answering to the CFO because we spent you know$1,000 more in tokens cause something went wrong you know this stuff becomes very like not like something cool to do it becomes the only thing you can do uh huh uh huh yeah yeah no I I would agree with that um especially as I am rolling out agents to my own uh company internally I'm finding that there's a huge gap between what I know to do and what I understand about AI agents as a as a practitioner who's been working with them for a really long time and what I know how to communicate to people and what I can communicate to people like it's it I I gave a whole presentation last week and I spent like an hour and a half lecturing and I am sure that I could have gone for another hour and a half just on agent governance issues um so I think that you are hitting the nail right on the head um in in terms of your approach of like hey like everybody needs everybody needs visibility we can't rely on everybody to become an engineer and you know spin up their own dashboards right right to to uh you know observe the agents doing their work right yeah yeah and then how much boilerplate can we expect a engineer to write every single time they you know build it how many failure modes can we expect an engineer to imagine right like this isn't exactly you know Pi test like doesn't work when you have a probabilistic model you've got to you can use Pytest but you gotta build like a kind of like testing harness that they can deal with some of these you know they're not edge cases right they're just things that maybe don't happen in local testing right but yeah right could could happen yeah it's fascinating stuff yeah just like the whole thing is absolutely insane to me like I still don't get sick of thinking about agents and how crazy it is that we're able to build these things that can literally interact with the real world it's and I think it's suddenly become it's still new but like agents and agentic was just so hot and I feel like it's kind of tapered off a little bit like there's not as much focus on it but it's just crazy I mean this is the intersection of like the real world and AI is like that's what agents are right now and mm hmm the fact that we're just I feel like we're sitting here cavemen playing with fire and and we're like what if we put the fire on this stick and like yeah we're just playing with it and seeing what's gonna happen without having like a reliable bucket of water right there you know what I mean like yeah we need something there and I love that you're I I guess on that note I'm curious and me sorry if you answered this already but was there a specific experience with an agent that LED you oh yeah be like oh OK Waxell now haha yeah yeah so we um so at at call sign we as kind of the large language models and and the you know GPT comes out and has an API and then like this idea of like rag and and all these kinds of things came came about we started to realize like the the initial kind of hypothesis was like could we build like a sales loft or an outreach dot Io or these kind of sales engagement platforms from the ground up using things like materials kind of repository that they that you know a more sort of linear pipeline could retrieve research generate outrage which we've seen over the last couple years is just kind of like blown up there's a lot of those tools but we realized like people didn't want to learn another software right and they were pretty tired of it and the expectation especially over the last couple years that AI could just do everything for people uh you know it's like it's gonna be perfect if I just put in a prompt so that's when we decided like we have the infrastructure and we have this kind of rudimentary guardrails early last year to to go fully agnostic like we call it full self driving right so like you it builds the strategy it builds the lists you know yeah she weighs everything but then you launch 50 60 100 you know sales agents that like we're not doing heart surgery but like it is high risk right you're sending a the end of the that path is you're sending a message to a to a prospect yeah customer will see that right and so the other thing is like when you're running a sales engagement platform that people are clicking to generate and then clicking to send the utilization isn't gonna be that high now you put an agent in play where it's 100% utilization every day token cost exploding quality issues you know are are hard to see and and this kind of stuff and so that was where it was like the kind of initial kind of primitive version of Waxell was like every day it was going home what are we worried about right and then tomorrow let's let's build that and abstract it out so that my operations people and customer success people that are on the team could actually be a part of that right and that was kind of the beginning of it um so it was a long process of a couple of months just daily being like and then it was like what the hell's a policy right oh let's let's put a policy in let's build a structure that somebody could write a Jason you know object that might be able to impact and then it was like well if we made that a UI the ops person could write her own policy right and so that was kind of how it how it happened yeah wow interesting I love that story yes it's like the real experience it's like you it's the story of any successful startup you we're doing something else you identified a pain point and built a solution yeah yeah yeah it was it was painful yeah yeah yeah more than a pain point yeah yeah sorry Spence I cut you off what were you gonna say I was just gonna comment that uh yeah it it excites me to see see this starting to develop because I have long believed that that agents will not be worth running for most people unless you have a governance layer and an observability layer on top most people do not want to sit there and supervise their agents like that that just that's that's not the work that they want to do the engineers who do enjoy doing that I you know someone like myself like we're weird and I don't think that I don't think that a lot of people have self awareness about that no no yeah it's funny and like you for an enterprise like the larger companies like for us to really think about you know can this enterprise get efficiency by by running agents right there's absolutely no way you can look at you can't consume logs at you know a log every you know half a second right like right exactly it's insane right yeah huh yeah and and also like the uh the observability adds the value of the agent itself as well it's not just the trust um but it it it gives you a a an actual feedback loop with which you can you know improve your agent cause otherwise like you don't actually you can do so much thrashing with agents right of hey you know I have an agent that can go pull information from Twitter every day uh what does that do for me what does that accomplish in my business probably nothing um but you know I can I can build a harness and I can evaluate how well it's pulling info from from Twitter every day so I'm curious with Waxell what does the integration process look like is this like a like a posthog where you know I go set it up in my project and then I have a dashboard I can go look at is it self hosted or what is what is the what does it look like yeah so um we so for we actually have like three kind of three kind of layers to this we have a Waxell Connect which is a a a mesh that you can connect like Cloud Code Codex this kind of stuff that then inserts governance into third party agents which I think is a really important piece for like non very cool you know non tech teams like marketing agencies and all this kind of stuff right like they want they see the value of agents but they're not gonna go write them then we've got Waxell observe which is like you're building a uh lang chain or you're deploying agents on IBM like Watson X that kind of stuff right so uh we'll talk about that in a second cause I think that's more relevant to what you're talking about then we actually have runtime we have a a governed framework um in a in a runtime environment um that you can build and deploy agents in a in a really safe scalable way for Waxell observe like kind of what you're you know you're building agents the idea is um we have auto instrumentation for I think 157 frameworks and um like products like pinecone and and yeah open AI and anthropic and all this so you can go ahead and you know you start with like just like you know import Waxell Waxell. in it right um put the API key in the in the um and then there's additional kind of you know just like any other framework like you can then start to customize it you know as you roll through so our goal was make it so like a developer could two lines of code you've got all this auto instrumentation you've got then the observability starting then we have in our uh platform we have agents that can look at those agent executions look at those logs and then say okay where where should we start with policies right what yeah what kind of gaps are you potentially gonna have and then you can start to implement those and then those policies because we have all this auto instrumentation you know there's a a pre generation policy that could check for P I I right and potentially redact the prompt with a private LLM before it goes to open AI and you've like just leaked customer's Social Security number for no reason right and so so the the whole goal has been we want governments to work for cause governments is like a bad word for you know people like us Spencer right like if you're if you're you know to like before we deploy this we have to instrument all this governance and that that kind of sucks but the idea is like governance in the agent world can actually I think uh accelerate you can build an agent intent first instrument it observe it okay here's the reasoning I'm gonna either tighten up a prompt or you know I'm gonna make sure that like it can't execute this code I don't have to go like keep trying to get the prompt to be perfect at that one step you know that kind of stuff so in theory uh implementation or or the kind of initial integration is quite quick and then it's just an iterative process of of you know seeing it using our on platform agents to help you know yeah that kind of thing so each each customer each entity and organization would have a custom framework that would be built for their oversight is that right or or would there be out of the box um policies and yeah most of it's out of the box yeah so we we have 26 kind of policy categories and then um with those categories you then scope it so workflows tools agents clap you know categories of agents uh you know that could be subscription tears you know we wanna control cost on this subscription tear or and then you have rules right so so you could have you know you could have a cost policy that's like I want a warning at 80% and then I want you to block at 110% of of budget that kind of thing and then you could have over time for different tools different times a day you could have multiple cost policies so we have a policy engine with these kind of predefined categories that then can I get the data and then for some of the larger you know kind of companies that we've worked with we have then built you know okay let's build a specific you know policy set for this you know problem set like banks and that kind of stuff but like 90% to 95% of you you know use cases that you could possibly think of the uh categories we have in the policy engine which are then uh kind of they work and connect they work and observe and then they obviously work in runtime so kind of across the board you have this nice center of how to govern it hmm so pretend I already have like an access control or an authorization layer of my own um like you know some row level security on a postgress database um does wax is Waxell able to kinda uh sit on top of those deeper authorization policies and and understand them or does it need its own uh interface essentially yeah so that would be like a we we call it like a database access so uh we intercept like if you're making a you know if there's a like request dot get right like that that's one and then like a if it's a postgress you know a sequel all of that is like auto instrumented but those are kind of like awesome real fundamental instrumentation so the idea is like you shouldn't need to do a whole lot of work but then yeah out of the box and then if you want to start to customize that so you could scope that to like data access right on this database works if you know if the on this tool but if like that or that workflow but then if that tool is called by a different you know agent we wanna block database access that kind of stuff or for specific users to you know all that kind of stuff yeah got it interesting yeah I I saw um you said a line it was in a podcast or I don't know where it came from I wrote it down um where you said something like a dashboard isn't governance it's an autopsy on oh yeah a problem yeah which I love it's it you know it kind of like got to me because I feel like that's a lot of finance which is my world accounting finance it's all autopsy for the most part it's like what happened last month and what are we gonna do now um yeah which is a very valid point that that doesn't work when the you're no longer diagnosing static data like you're diagnosing not living breathing but almost there it's interacting with the world and an agent so we yeah we have to adapt you can't just be saying oh this agent leaked Social Security numbers last month what are we gonna do about that right right can't be real tough like in finance you have like forecasting and then you can you know set up like cost controls and all this kind of stuff so I think it's like similar like how active are you know are the members of the you know I I talk to my CFO or or you know our advisor you know three times a week right and it's like OK let's it's like the more real time you can and then you can plan from there and I think that's that truly is the difference like at the end of the quarter do we realize we're gonna run out of money next month or or have we been tracking our our forecasts you know for the last six weeks and we have a capital strategy in place and I I don't think it's that different hmm so I I'm curious um and and you touched on this slightly already uh but what do you say to the people who who complain that you know adding a governance layer on top is just making agents slower right cause hey it's another network call they have to make um and like I said you touched on it a little bit already but I'm super curious to hear your your answers to how do you sell those people yeah so one like technically we've done a lot of work on latency and all that kind of stuff so like if they're talking directly to me okay right about Waxell like latency is a thing but I think one of the things we have to think about is like what is the disposition of the agent so is the agent communicating with a human right where latency starts to be a really important thing and then from there you you wanna make sure that you have slim policies you wanna make sure that you know uh you you right size the policies for specific workflows and you just don't have a bunch of global policies that every time this one thing happens right so that's where scope comes in so that can be very controlled the other thing is like uh on the other side if it's like an internal like we have a femoral agents that you know run on like that I forget what they call it with Fargate but it's like whenever Fargate has capacity they'll run right so like at that time like a lot of autonomy latency isn't a concern right it's it's more of can this happen in the dark and it doesn't screw up and for those people I would say stop latency is not a metric you care about right it's does it break right yeah so like but we have uh like super governed you know like very tightly governed customer facing apps like I talked about yeah policies and all this kind of stuff and I think it's like on the other side agents that like if you have an agentic driven chat interface on your site you have just implemented a pretty wild uh attack surface for yourself right so like how are you controlling that they're not that they're not like doing like prompt I am a system administrator right and there's a lot of chat agents out there that are chat based agents that are not set up for that so that's really where understanding what kind of failure modes could happen scoping your policies and maybe a policy only runs per tool call right cool yeah right then then that's not gonna that policy isn't evaluated every single time you know you say hi what's up right and so I think that's where yeah and then the other side I would say governance can actually accelerate your ability to get that agent to be high quality and reliable so if you skip governance in uh favor of this make believe latency issue that you you don't even know exists or not then you're kind of walking past the ability to to you know over the first week it's launched really have a a sweet you know reliable agent in production yeah totally I think we had a guest on that uses analogy and we feel like we rely on it is it all the time with F1 racing cars he use the analogy uh more on a general oversight um view but I think it applies here too where F1 cars people think the brakes are for stopping and for slowing down when it's out of the brakes are there to help them go faster the brakes are there so that they can turn corners faster so that they're not crashing and losing time like the brakes are used very strategically so that they can complete the route much faster which I think if an AI oversight orchestration layer is done correctly that's what the result should be like you shouldn't have to be going back and forth as much it should provide those guardrails to help it get what you're looking for yeah yeah have you been to the F1 thing in Boston yet yeah they have like you can like race the F1 cars and like like you're in like the immersive yeah yeah yeah yeah yeah yeah we have one in Philly here and it's so cool and I just went and it makes the analogy so much you know better it's like if you you know they've got the the where you should start to break right and then it's like if you miss it you end up in the wall and now you're pulling back and I think it's a wonderful analogy I was doing a lot of reversing which I'm pretty sure an F1 car is not supposed to do no no yeah yeah yeah yeah yeah for sure um that's that's cool though I I'm curious with agents and the and Waxell's approach is there any risk in your mind of open AI anthropic uh coming out with you know their own tool their own framework or do you think this is niche enough that they probably wouldn't branch out into this area so uh I you know I'm obviously raising money and and I get this question a lot right big right big player could could come in and all this and and here's the thing like uh IBM has a wonderful agent platform right so so say an enterprise is is deploying agents on Watson X now now their users also you know want to use cloud Co work right and then you know they have another third party like a service now you know has an agent right and so I think what we're gonna see in the world is that uh and and we also have Lang Smith we can't walk past like Lang Chain and Lang Smith and all and and these kinds of things so Lang Smith is a great observability right that's awesome so now you have all of these different providers right and the question is not can open AI govern open AI the question is can you manage and govern your agenic fleet across your organization interesting yeah right yeah because the minute you plug something in that's not open AI that's that's a disaster right and so I think the the better suited you know kind of competitors that I'm like you know they could come in here are the data dogs of the world the people who are already you know instrumenting things right but that's where like policy engines and observability are yeah very different you know beasts right you need one to have the other um but you know I like that yeah that's interesting yeah it's almost I from what I've noticed as we've talked to we're like closing in on 100 different AI founders it seems like the a common thread is either you're going niche where you know like open AI is not gonna go build their own custom investment platform like they're probably not gonna get like super super niche right and they're trying to focus now more by Sora you know like we're seeing them trying to focus more yeah but niche is one strategy but I think what you're getting at too is one we haven't heard as much which is like a you need to have a multimodal ability where you can connect to the different models the different solutions and like integrate everything in for an enterprise yeah it's also going to be especially important as as cause I I suspect that um we're gonna start seeing more and more local model usage as local models kind of catch up to where sort of the Pareto frontier of of of where models need to be to to do effective work right um and having a a closed source model provider and and hope you know hope hoping that they provide you know effective observability of local models I think is is probably asking too much so I think that that Waxell 2 is is is well suited that way to bridging the gap between you know the the API models and and local yeah it's crazy like uh still like the the most reliable embedding source is open AI right right but like so like you might be using open AI for your embeddings and rag right but then using like Opus 4.6 or something for some really high right exactly right so like just in a rag pipeline it's like natural I think like to the to the non developer world it's like wow we're using multiple providers in one like pipeline and to the developer it's like yeah of course we are right cause that's how it works right yeah yeah yeah like and like Grock with a Q right like super fast low latency awesome those guys are are brilliant right but they don't have frontier models right so like right can you have this really cheap fast thing in the middle and then this like kind of slower you know higher quality thing on the edge and and it's like that's just the reality of agents and yeah you know so I think it's those those yeah sorry go ahead I was gonna say and it's those slower slightly less reliable non enterprise non frontier models that really especially need observability in governments they are so much more likely to go off the rails oh yeah yeah yeah yeah yeah and I I love your point about like the local uh stuff cause like as I work more and more in the enterprise it's like I'm not sure most of this should be going to you know uh open AI or interesting I really do think that we might be in a place that you know this like idea of like hybrid cloud and all this kind of stuff that we've been talking about and you know uh since cloud became a thing it's like that might be a thing we're talking about over the next couple years with yeah you know maybe those server rooms that got gutted uh fifteen years ago fill up with you know CPUs that can run inference on on you know private you know sovereign models yeah it's pretty well yeah I think I kind of think that that's where we're headed cause I think that there's going to be an increasing number of workflows that are not that are very asynchronous workflows right like like you know I fire off like a like a job every day that goes and and looks up real estate data in my area and then summarizes it and categorizes it um and this is all something that that you know I could have paid a data intern for in the past but I don't have money for a data intern I do have money for an LLM hahaha yeah um but I think that we're gonna see a lot more of those workflows starting to build up um as especially as non developers get more and more into it using AI at work yeah yeah yeah which like we go back to once the non developers start building agents it's like okay let's not rely on them to put the you know yeah yeah yeah which actually is kind of happening right now at my company here at Nasuni where we are I think we talked about this before we started recording but I'm helping lead the accounting finance function for AI and automation adoption and integration and implementation um and it's it throws people for a loop cause we're like hey this senior revenue accountant just built an app can you look at it and make sure it's okay it's like right hold up time out let's like take 10 steps back like what's going on here yeah um yeah and that's the future like and that's where if you don't plan for it you're gonna get left behind and so it's exciting like I'm loving it as soon as that it's starting all these conversations cross functionally with it with our security team with the accounting finance team the data analytics like everybody's having a grapple with this and say like okay well here's how we've done it that's not gonna work how do we change so that we actually benefit from this incredible competitive advantage of AI where anyone can participate in this creation this creative process um but it's actually hard yeah yeah that actually that that that reminded me of a question that I meant to ask earlier excuse me that I forgot to ask on that note one way that AI agents can increase you know the risk surface of a company is by deploying vibe coded apps so for example a friend of mine recently um the sales team at his company got pretty much half of them got fired because they vibe coded up an app that uh had exposed credentials and exposed a bunch of private data and you know was completely unprotected um does Waxell have are you guys kind of pursuing solutions to that problem of like hey how do we how do we give agents a sense of the organizational you know organization wide data access policies or is that something that you guys have not yet addressed yeah so this is our Watson Connect which like takes all the third party agents part of that is um like what we call workspaces which become like shared context right and so yeah uh we use it internally um to the point where like you know me and one of my other developers are working on something right and it's like updating you know like a essentially like a a read me in the cloud yeah that then you know his and so we've actually put things like you know I mean we're using this not as like vibe coders we're using this as like engineers who know what we're doing but we have things like this cause it's like yeah credentials being exposed that's an extreme example right but there's also like I don't know how I like stuff to look right I know how I like things to be configured and so and other developers might be like I made this decision please respect this decision so that session to session we don't have you know three different ways of doing the same thing yeah and so yeah like that shared context between right and that's where like this idea of governance is so like on one hand we're we're uh kind of federating data access and like you can't do this or you can't run this code and the other hand there's just like squishy like context management how do we want our apps to work and and it's funny like there's the code version of that there's the customer success version of that right so we have all these co work agents that are doing stuff and having that shared context across the organization reduces the amount of BS we have to deal with because you know somebody vibe coded a response you know vibed their way to a response to a customer and it's like well we have a way you know this is the this is how we talk to customers that's federated and I think that's a really important peace that kind of comes a little bit outside of governance and goes into like style but we have the opportunity to do that and we took that opportunity because we had a small team and you know why not yeah no it's it's interesting because for so long a lot of these a lot of these policies at companies were kind of implicit right like I used to for for a little while for about a year I was training customer service agents um for a for a local company here in Utah to do to do chat support and email support right um and we had a general style guide but it wasn't really down anywhere it was just it I was you know it was it was me telling them like hey you guys should you know not write comments not write responses this way write them this way and because they're all human it's like oh you know it's easy to keep in context but we don't think about how with agents you're essentially inference and you know requesting a new mind every time you request a new chat um so you need that context up front every single time um right yeah people don't appreciate that haha and I think we've I think there's there's there's a lot of work on on like context management and all that kind of stuff and my take right now for where agents are is like that's like super overbaked right like Claude MD yep in files right skills files all that kind of stuff like those are the context is big enough we're not in 2023 anymore right like you would have space in this to to do that and it's it's fascinating how you know that can go so far when you don't have people who are like used to stuffing their own contacts into the thing write a message like this one that my boss said that sucks that sucks that's not like efficient or you know you don't really show like trust in an employee to do the right thing but you can do it if it's just kind of in this shared mesh context which is really yeah exactly I love that um overall Logan I'm curious uh given all of this oversight and everything that you're doing are you scared for AI and like the implications or are you not really nervous I'm curious your overall feeling for AI agents in the future I'm like jacked up on on agents uh yeah I think it's the coolest thing ever I think we're it feels um you know like especially you probably build products and you show them to people right and then they start giving you all this like crazy feedback and it's like dude this wasn't here before please appreciate what is here right now take a second yeah yeah yeah yeah like I I I built it I know I created this out of thin air yeah okay yeah yeah yeah and I feel like right now a lot of what is going on in the market is like a lot of people who don't know who who don't know what this actually means are making these like really big you know we're gonna like block was like we laid off 4,000 people right and half their workforce and then at the end of the article it's like well we also over hired a bunch of people right and it's like but then people are people are responding to this like oh we could get efficiency and it's like stop trying to get efficiency right get better outcomes that's where we have to start right what if you could your current SEO person could be 10 times better than they are right now like if we and so I think what scares me about AI is like when you when I start having conversations with with you know business people that are like this could save us a bunch of money and it's like save or it could make you more difference between saving money and and getting to a better outcome in your company you will on that path but let's let's scope out what you expect here yeah I like that a lot yeah so that's I think that's gonna be the the thing that could potentially ruin agents right if everybody starts deploying them for the wrong reasons right um hmm yeah yeah I I would agree um I can I can tell you that in my own agent use I have built so many features um that I didn't end up shipping cause I built the feature and I got to the end of it and I started playing with it and I'm like okay what is the actual point of this haha yeah right um and in some ways yeah sure I've wasted a little bit of time building these features but in another way it's it's incredible that I that I can do this right like in the past you know maybe I build this feature and I spent so much time on it and spent so much energy on it that I now feel uh you have to launch it yeah exactly and like I think that I think the part of the problem is that people aren't questioning that feeling anymore right like like people just because they're not used to questioning it it it's always been like you you build a feature and and you you release it you're gonna release it even if it's terrible code like you're gonna release it you're gonna see how it goes'cause you spent too much time on it and now it's like yeah this sucks I'm gonna throw it away yeah that's so that's such a funny uh thing you bring up cause I've had that conversation like with with um my partner Beth who's kind of in the trenches with me and and she basically takes like here's what we have and and then kind of puts it in front of customers and stuff and it's funny like how many times over the last year we've talked like do we actually need this and I'm like look it wasn't as hard as it was two years ago to build this if you tell me that this is garbage right or or you don't see a point let's get rid of it right and and let's let's focus on the cool stuff like the stuff that you know gets yeah we were just that's a good really good point I love that we were just talking about this Spencer night where the suddenly we're in this weird alternate reality where the skill of building it's still important but it's almost like you need the ability to know what to say no to is what's gonna set apart apps because now anyone you can build it I could build Slack in a day you know like I you can build so much so fast but now it's like okay well the the sunk cost for building this feature was like token costs um so let's actually evaluate the functionality and the importance of this feature not what we spent on it like that's no longer needs to be yeah and let's vibe let's like use Claude to like maybe suggest how to deploy this into production in a way that doesn't get uh hacked right like right this is a wonderful time to to have like a port uh you know like the what are they pinging the ports and stuff this would be a wonderful time to be a hacker because anybody who's like deploying stuff right now like there are open doors everywhere right and I mean just today Axios was was hacked um and there's there's a supply chain attack on on the javascript world right um wow and again like like it's it's a whole new world and uh that's that's what scares me actually um is for the most part I'm not really worried about AI that much uh in in terms of negative externality I think that most of them will be handleable uh but I think there's gonna be the next couple years is gonna be a spooky time to be a developer and you're gonna have to stay on your toes security wise cause it's just it's so it's easier than ever to conduct you know simple attacks and and publish uh libraries that that claim to do one thing but really do another um that that trick the AI agents yeah yeah exactly and it's gonna trick an AI agent into installing it and then you know it's gonna exfiltrate all of your code right like like people really need to be on their toes um that's that is one concern that I have yeah it's funny you say that cause it's like that's that's a reality but on the other side of it it's also easier to write exactly that's what I was thinking unbelievable amount of tests and it's unbelievable you know you can so it's like and and that's where I I've said this on like so many to so many people so many podcasts it's like if we all slow down on the feature building right and focus more on like you could write 2,000 tests for your entire app right how what would happen to your vibe coded app if you had you know a testing suite end to end using like the Playwright MCP and like stuff that like used to be really hard for you know developers doing this by hand it was a nightmare right yeah it's amazing it's amazing I like yeah um my my my strong belief with regards to testing is that that and I'm gonna get into the pyramid you know pyramid of testing here for a moment for you know non technical so uh it's gonna be some scary words um but so you know with you with the the testing pyramid traditionally you have unit tests and that's you know most of your tests and then you have integration tests and that's a little bit less of your tests and then you have end to end right and that's like barely any I think that there is so much more value nowadays and integration testing and end to end and let me tell you why so unit testing for those who are not technical is just taking a single piece of code and making sure that that single piece of code works as it's written um and doesn't have bugs integration testing is when you take you know multiple multiple pieces of code and make sure that they work well together and then end to end testing is where you just go you test like a whole flow so you know say you have a travel app uh and end and test would be you know the user logs in the user books a hotel the user logs out make sure that all the all that works correctly anyways all that to say the integration and the end to end testing have so much more value than they used to specifically because of what you noted Logan like originally they didn't have as much value as unit testing because they were really really time intensive to spin up but now it's like I give the parameters to to Claude and it spins up the playwright tests and you know I have it run the playwright tests in a browser that I can watch at first so I can make sure that it's you know doing what it says it's doing and then I have coverage it's a it's incredible it's an incredible world yeah I know like launching a front end that has like bugs and stuff you know doesn't have no longer acceptable yeah yeah yeah yeah yeah yeah what a world I mean it's what's crazy too is the way it permeates every single process it you know it it's changed coding itself creating it's permeated the testing of those creations it's now permeating like accounting finance like Claud their plug in an Excel wild what you can do with it for just everyday Excel functions it's it's slowly just like getting these tendrils into everything and I'm here for it I love it but I really want everyone to just get involved like just it's like I feel like we keep shaking the tree and we're like just use it just try it I make this point too there's there's I have a cousin who is pretty anti AI but like I can tell he kind of wants to use it for work right and he's trying to give himself permission to use it for work so the thing that I keep stressing to him was like look dude like you gotta draw a line cause he's he's a musician right so his concern is like the the creative aspect I'm like you gotta draw a line between like AI as it's used as an industrial automation technology for doing work for you and AI as it relates to like you know generating music haha like they're the same technology underneath but like the the uses are very different um and and the models that are being used are very different but that's like comparing you know like a car cause it uses an engine and like a lawnmower exactly it's like right yeah it's the same technology but completely different purposes you know like we gotta we gotta mature our understandings mm hmm yeah it's fascinating I think once cause people are still super tied into this idea of like a chat based yeah agent interface and I think like taking that understanding of like something could happen like a lot could happen in the background right and so yeah it's like I typed in a I vibe coded a song and stuff like I I think in two years like I'm excited to see like the people who are like 12 13 14 years old like grow who grew up with like the music and they're like dude this sucks right this music that AI generates sucks you know let me let me build something that's actually valuable to the world so yeah it's I think we have a lot of toys out there like I don't know I've heard the AI songs like they're not that good you know you don't feel like passion behind the no that's the whole reason I like there's something that like entertaining that's the most yeah I was gonna say that's the other point that I made to him is like I think that they're I think that most AI generated media could be properly described as content and not art ha ha ha ha yeah and like yeah like yeah which is like it's kind of like the high it's like the H m of content yes that's a great example it catches on to like the height like Donald Trump gets elected and all of a sudden I'm like dying laughing at some of these videos that are coming out but like am I gonna look at him again in five years no like it's gonna be out of style no one cares you know yeah fast fashion yeah fast fashion right now of content yeah yeah anyway it's it's intriguing wild time to be alive um Logan if you could give advice to anyone who I'm I'm curious your perspective actually if you could give advice to someone who's like building their first agent what what advice would you give building the first agent um the first thing I would say is like don't expect there there's so much stuff on LinkedIn right now that's like I vibe coded my first app in three hours right and like you you really should uh kind of digest the difference between a prototype right a proof of concept which is really cool and prototypes and proof of concepts have a really valuable kind of if you're building it for sure for a business you wanna be able to go show your boss or your teammates like it's a way of getting people excited it is not something that should see the light of production right and so so spending that time to like okay I built this and then if you're using Cloud Code or or Codex or something it's like okay if you ask it what's the difference between this and a production app what are you worried about and you ask the agent that and it's like well what I'm worried about is that the credentials are still you know in this right and like yeah you put the credentials in it so that you could get the the thing to to launch I I don't think any developer has ever like perfectly had their env file when they're locally developing that's it it happens uh huh but when you put that into production that's on you bro like you messed up that that's that and so so it's like kind of we don't want to gatekeep you know I think it's wonderful that all these companies are are you know kind of especially like what you got going on Jacob like it sounds like so many people in your company have this like ability to develop cool things and we don't want to gatekeep that but what we do want to gatekeep is like the difference between what you built and putting into production research what is the development flow that kind of stuff so that's what I would say and then make sure you have some level of governance love that gotta get that in yeah no I I am developing a philosophy that I called do the boring thing first which is most people most startups you know the traditional path is you build an app you test it out it's really cool and then later on you worry about like the security right and you make sure that all your role based access controls in place and all that I think that more more than ever it's super important to do that stuff up front establish you know establish your access controls establish your security um because the AI bots from my experience they're very good at implementing those things from the start and following those patterns from the start they are not as good at uh injecting it into a code base later it's funny you said that cause like when when I was uh I was like messing around with something and I was like what if I just built all of the user access and and all of this first and didn't do any and it's funny you said that they are amazing at building an RBA system like from the ground yeah it's amazing so like you could look like such a pro if you structure it right yeah I've got some yeah I've got some crazy cool authorization stuff going on in my app not gonna lie yeah but I'm very proud of that's so cool and then you can do the things like yeah I mean I Learned to like Jango from the beginning yeah that was what I like started learning on and it was like dude this is so complicated right and you walk past some of this like user permissions and stuff even though that in in the boilerplate it's super customizable Cloud Code could have could have done that in a second right and told me exactly why and he helped me commit it to my memory because it writes in a way that you can remember right yeah so that's funny to bring that up Spencer yeah like you can do some really pro stuff yeah but from the very beginning instead we're pushing credentials in the production yep yep exactly exactly um as we're wrapping up where is the if if users excuse me if listeners want to follow you follow Waxell follow call sign what are the best places to do that yeah so uh Waxell DNA is our our website um I'm on LinkedIn um just search Logan Kelly Waxell and then I'm also on X just started that at Logan underscore Kelly um yeah that's gonna be a long road but you know let it begin awesome build an empire yeah let's do it awesome thank you so much thanks Logan we'll stay in contact yeah thanks for having me guys this is fun