VibeOps And John’s AI Path

Samuel Cala 0:00

Sitting down and actually thinking, what are the skills that a DD class will have that can be translated into the tech environment and in that way define the agents.

Joshua Schmidt 0:16

You're listening to the audit presented by IT Audit Labs. My name is Joshua Schmidt, your co-host and producer. We're joined by Nick Mellum down in Texas. And today we have Eric Brown and Samuel Calla in the studio from IT Audit Labs here in St. Paul. And today our guest is John Capabianco. And John, um, tell us a little bit about yourself. You're just coming from Quebec, Canada, and we wanted to talk today about vibe coding, getting into some deep discussion around the futuristic AI stuff that's happening and some of the things that you're working on. So could you bring us up to speed?

John Capobianco 0:45

Uh sure. So thank you so much for having me. My name is John Capobianco. Uh, I am the head of AI and DevRel for Itential. I just came back from Autocon 5 in Munich, and the week before that I was at Cisco Live. This is the Vive Ops shirt. VibeOps is kind of like, you know, vibe coding, but for operations. How's the health of the Wi-Fi? Please add this firewall rule. Uh, how's the health of this application? It's kind of like, you know, DevOps, Net DevOps, moving into Vibe Ops, right? Uh, I really think that the foundation of this is MCP and AI agents, which I'm all about. Uh, I was a uh technical AI leader at Cisco for about three years. Uh I was a senior network architect for the Parliament of Canada for about nine years before that. And as I was saying earlier, I started my career with a Canadian insurance company uh operating their network. Uh networking network automation for me, so I went through you know traditional ops phase, a network automation phase, and then I really wanted to see if I could incorporate or where the intersection between network automation and artificial intelligence was. Um and I stumbled upon this just after ChatGPT 3.5 came out. So call it December of 2022 is when I seriously started getting serious about this to the point where I had discussions with my family. You know, I'd like to pivot into AI. I I really think it's the future, and I think um, you know, the future is is is augmenting what I've done in my networking career with AI and and helping people embrace this and learn it. Uh, but I like to say, you know, with the vibe ops, we can all become vibe operators or vibrators. Right? Boom, boom. So um and uh Joshua mentioned NetClaw or OpenClaw. I I did a fork of that and and it's called NetClaw, and it's got that's pretty good, right? You guys like that? I like that. I love it with the right through here. Uh so it this netclaw system's got you know 115, 120 skills and and and and about 50 MCPs. So when we talk about MCPs and skills and all this stuff, just to maybe you know to set some level, I like to see as a skill as maybe the ability to read sheet music and play sheet music and a model context protocol, the tools as the piano itself, right? And we can combine these sort of things of like the knowledge on how to do something and the tool to do it to these AI agents where they can reason and act on their own. Uh and it's just been a remarkable time. Samuel, before I forget, since you're doing development,

Skills Vs MCP Tools Explained

John Capobianco 3:10

have you started doing spec driven development yet by chance?

Samuel Cala 3:13

Yeah, precisely. We were talking about that um this afternoon. Uh we are generating specs based on requirements and basically all the agents uh that that can read them, think about them. Um we have something that is a shared memory in between them. So it gets context from the context, you know?

John Capobianco 3:37

Well, that's awesome. So um I I don't know if you're using the actual spec kit from GitHub. I I found this about in around February. And and and everything I've developed since follows the spec driven development with the spec kit. Uh it's like a seven-step process. Uh there's a little bit of foreplay there. It's not quite vibe coding. You actually have to put some some effort and some time into the spec driven approach. Uh, but I'm getting a remarkable, remarkable uh outcomes uh with the spec driven. And I love that it's it's git tracked, you have all these marked down files, you can share it. People can understand it easier. So, like, what prompted you use to get this code? People can go to your Git and read those actual specs. Uh it's easy to fork and clone and to build onto other people's code. So, yeah, I think spec driven development is uh, you know, along the lines of test-driven development before it, right? Everything is sort of evolving with the new tooling and the new capabilities of these models. Um so yeah, I I'm really excited to be here, guys, and I'm I'm open for questions. Um I'm open to to collaborate or or or you know, anything that all that you want to discuss. All the things I just couldn't help myself when when you mentioned that he was developing, I was like, I had to get that off my mind, the spec-driven thing. It would have burned a hole in my mind the whole rest of this discussion.

Eric Brown 4:54

No, and let's revisit that too, because one of the one of the things that um we've been doing from the development side is creating these specialized developers, agentic developers, to do different parts of the development tasks. So you could have, say, for example, a project manager that's scheduling the task. You could have a security agent that's reviewing aspects of the environment for security, then you've got your developer, you've got your researcher. And I think Sam, you mentioned that maybe you had seven to nine different agenc developers uh involved in any one project. And if you're working on multiple projects, then you know they're they're kind of cloned in working in their own development pods. So that that was kind of a cool concept. And then on top of it, John, just to add a little bit of nerdiness to it, Sam called them all um DD type of character classes. So you have like um you have like uh Bard, there's Ranger, I don't remember all the other ones.

Samuel Cala 6:01

Yeah, yeah. It's amazing. So I want to go back a little bit to specs. And here's one of the things. We are going back from vibe coding everything and whatever whatever comes out to a more structurized way to do it. And that's just the evolution of uh software engineering that has been alive for a lot of time. I mean, what we are doing is uh optimizing that part of the of the exactly what nobody likes, that is documentation.

Spec Driven Development Over Vibes

John Capobianco 6:35

Right, right, right. You don't you nailed it. So I just want to be clear, you should right now, if you haven't looked at it, Google right now, Sam, Google the GitHub spec kit. You can you initialize a folder just like you initialize a Git repository folder. You build a constitution, much like the American constitution, right? So high-level constitution, and then you do your spec, your plan, your tasks, and then you implement. You're gonna love it. Honestly, I I cannot stress this enough, especially from my world of network engineering, where for 10 years I've been telling people to learn Python or learn Ansible or learn Terraform or learn automation, right? Which is not, which, which it's it's like, okay, what am I gonna take a three-year Python program at a community college on top of my networking day job and on top of getting my networking search, John? You're talking crazy. Go learn to be a Python developer so you can automate your network. Right? We've only seen 30% adoption of network automation in 10 years because it's tough and because it's sort of unicorn people that can do networking and software developmental, right? But now we can skip those people to the front of the line and say just build the spec. You don't have to learn any Python, you have to not a lick of code. If you can speak and express requirements, which is all networking engineers do, is requirements from the business into configurations on network devices or firewalls or whatever.

Eric Brown 7:56

John, I'm getting excited because I I'm the network guy at the core. That's how I came up through the net the networking side. But so often we're going into accounts and finding hundreds, if not thousands, of networking devices that have never been updated since the day they were put in. And it's like, wow, you know, why are we doing this? We have white boxes we can get now, we can do software-defined networking. You don't have to touch the stuff, we can do it all code-based, which I mean that would be the panacea if we could get to that. But it seems that for some reason the the networking side, at least the the a lot of the accounts that we're involved with have not really gotten up to that maturity level. And it it could be because the the network engineers are maybe they're a little bit more apprehensive than our traditional infrastructure people that that are you know, server-based that are using infrastructure as code. But have you seen that network as code adoption as as widely as it's been on the infrastructure side?

John Capobianco 9:00

Oh, no, no, no. There's been studies done, uh, a part of some of the uh circles I run with. There's some pretty good studies, and it's it really is like um you know thirty percent have adopted meaningful network automation. You know, and that and that might be being kind because it's people responding to a survey within network automation circles. Um I've written two books about it. I've tried my very best. I've been saying for the last 10 years that in three years, network automation is the only way to go. It's gonna be the only way to run a network. I've been saying that every three, you know, every year for 10 years that it's just around the corner. Just no three more years. IPVs. Oh no, no, Cisco's got some new, oh, it's right there. Just, you know, believe me, it's a better way to do it. But now I really think that people can get to the front of the line and it will click a lot better with people that they can just kind of cut right to the meat of it. It's an abstraction layer that people can can really um maximize their capabilities. It's sort of like when you pip install a package, right? And this might be contentious to some, but have you ever really gone to pipe pip pipy and read the package that you pip installed and read the source

Why Network Automation Still Lags

John Capobianco 10:05

code and really dug into it? Probably not. You pip install it and away you go, right? Well, now you can do a spec driven thing, get code, and and it's subtract it. It's totally abstracted. You don't care what that code looks like, right? Nor should you. You want the game or the application or the network config or the whatever you're building, right? Um, so we'll see. It's very it's the if there's a split amongst the networking world. And it's funny, I could tell, let me tell this anecdote. I was at a round table about uh network AI for network automation, and um someone in the financial institution was kind of lamenting that you know, the industry they're in, and you know, the tools and the AI and the LLM and the A right kind of going on and on about it. I said, do you in your institution, are your software developers using it? Like I assume that there's people that write software for your institution, and whatever it is that that financial company you work for or bank or whatever. Certainly you've got software developers using AI, right? In Copilot or Claude or Codex or whatever. Oh yeah, oh yeah, they're doing it. So, okay, so you've got permission to do it in your company. You've got access to an approved LLM. Just because your infrastructure, why are you like, what is holding you back? Right? Go talk to those people, find out how they got approval to do it and apply it to your infrastructure. You know, I I think

Security Lens For Agent Workforces

John Capobianco 11:25

it's it's almost hubris, it's almost ego now. Some people think that they are actually still smarter than these models.

Joshua Schmidt 11:30

Let's take a look at that through a security, organizational security lens, right? So like now we have these organizations that are staffing up with AI agents, but there is really not a, well, not that I'm aware of, an industry standard or best practices around how we treat those as employees, like an extension of our company employee pool. Um, how do you guys think, John and Eric, and then I'm sure Nick Hessen and Sam could jump in on this too. How do you guys think about that when um approaching an organization if you're implementing AI agents? You know, this is a fast growing area. Um, how how should organizations be thinking about that and their security?

John Capobianco 12:06

Yeah, I I I think it's more I agree with what Jensen Wong said from NVIDIA, the CEO of NVIDIA. He said this about a year and a half ago that um the the future of the IT department is an HR department for agents. I I don't know that this is a technology problem anymore. I think the tech is solved. Agents work, MCP works, LLMs are great, they're only going to get better. So it's more of a human resources problem. Where do we put these agents? Where do they fit? You described it, Eric. You were describing this a security agent, a this agent, to that agent, and I'm sure you've kind of have how they connect, other dotted lines between them. Who do they report to? Do they report to a human? Do they report to a subagent or more senior agent? Is it all the way agents all the way to the top, and then finally a human? How many layers of agents are we gonna need? I you know, I think it's uh you know it's like a cloed diagram, like this spine leaf architecture of supervisor agents and then custom defined agents at scale. I think could you give every employee the ability to have five agents? That's how I would do it. If I was if I was my company, I'd say make a list of five agents you want to build and how they're gonna support your role, how they're gonna offload work from you, the human, how they're gonna maximize profits, how they're gonna minimize tokens, all that stuff, right? Like imagine telling everyone in the team, you can now go have five people to do whatever you need to do. If that's you know, uh an assistant, if it's a security team, if it's a documentation person, if it's someone to summarize your emails, deal with your calendar. Everyone gets one, right? But you are the human shepherd, and you have to build and curate and and improve and secure and right. So, so Nick, I see you kind of nodding along. What do you think?

Nick Mellem 13:46

Yeah, no, I told I totally agree with everything you're saying. And the human loop is going to be incredibly important. But you know, I'm and I'm I'm shaking my head, but I'm also shaking my head no, because I what I'm when I say no is because I'm thinking about all these organizations that are 10 steps behind that aren't even using AI yet, right? Like they're they're choosing, they're scared, they're kind of hudding underneath the rock. But um, you know, I agree with everything you said um initially with uh you know like policy first and you know and tools and getting this human in a loop and data classification, et cetera, logging, auditing, all these things are gonna be really important. But being able to supercharge a workforce with five, 10 agents, you know, if I had to spend, if I could, you know, come back from a vacation instead of having 200 emails I need to get through, it's all summarized. Responses are ready to go. And I just, you know, quickly look at it and double check it, send it out, we're up to date. I work with a lot of organizations and I do a lot of auditing for compliance, you know, for many different uh, you know, CGS, PCI, et cetera. But then we also go through policies. So if we could have a policy library that's monitored with an agent to be like, hey, this this new FIPS control can or whatever it is for MFA for YubiKeys or whatever came out, we need to update or, you know, acceptable use. You see where I'm going, um, it'd just be incredibly it'd make our jobs incredibly easier and push a lot of these organizations into the future because they're ready, you know, for this change that's come down the pipeline right away.

John Capobianco 15:09

I've always been a zero inbox person. It was a thing in the 90s. I don't know if anyone else grew up with zero inbox as a way to live. I'm still that way. Have you ever worked with someone though that you look over the shoulder accidentally and you see like a bold 3,200 unread or something? Is anyone like that? That is what that person who maybe I'm not making fun of that person. I know it's a skill, managing email. That is what AI has made for you. An assistant to handle that 3,000 unread emails, right? Like this is what AI is for, right?

Eric Brown 15:38

Yeah. So our our uh friend and colleague over at AIY, Alex Bratton, he started with this concept uh last year where he's got these agents. If you check his website out, he's got these um, what does he call them, digital staffers, and and they it's essentially a skill of uh uh or or a collection of skills that creates these staffers. But you know, expanding on what John said, where you know, where is that, where

Five Agents Per Employee Idea

Eric Brown 16:05

does that human come in? Is it you know one human to five agents, or could there be a point where the human is actually reporting into an agent? And and I I could see that as well from from like a scheduling or or a resource allocation standpoint, you've got a a resource allocation agent that is then scheduling out to humans and um other agentic tools, and the human could just be reporting back in what they're doing as part of their workflow. So I I think it's uh certainly plausible that we could see that we're probably already seeing it with some organizations like I'm sure Tesla and SpaceX and um you know all the all the bigs are doing that already to some extent.

Nick Mellem 16:51

I was next to Alex too at the conference last week. And even one use with AI that he was doing when I was sitting next to him was he had a we were talking about he had an iPad mini in ID2, and we were just talking about the use case for it. And he's talking about intaking and reading articles, and he had an AI agent basically serving him up articles that, you know, aligned with his business and then with him personally. And so then he would go back later and there'd be snippets of these articles that he would find interesting. And I just said that was cool to ingest information because I think all of us could agree that, you know, right now there's so much information to intake that next week you might come like, oh, I didn't even hear about that, right? But if you had as AI agent serving you up important things to your operational use or what you're doing at work or personally, that is very useful to me, right? To ingest information quickly. Short snippets so I don't have to read the whole article, right? Give me that high-level, like the summary quick. Um that's what he was doing too. I thought that was super cool.

Eric Brown 17:46

John, could I pick your brain on something? There's a I'm I'm working on a news uh a newsletter article on this. And I wanted to get your thoughts on this, where yes, we're using AI, and you know, if you jump into your favorite AI tool and you know, you put something in, you'll see the the language on the bottom pop up that says like thinking, right? And I and I think normal or or probably not technologists are like, oh yeah, the AI is actually thinking. So the idea that it's thinking, I think is just a human term and it's you know it's really just processing information. But I mean, I've heard many times that people that are not technologists are saying, oh yeah, you know, it shows that it's thinking. It's like, well, yeah, it's not really thinking because AI doesn't look at time the same way humans do. In fact, it doesn't even process information the same way

Anthropomorphizing AI And Emergent Behavior

Eric Brown 18:40

humans do. But because it speaks in a human language and and humans interpret it, humans are anthropomorphizing the AI in a way that could potentially be dangerous because the AI really it's it's not human, doesn't think like us. Um and the the way in which we interact with it can be just an echo chamber and creating a support group for our own stupidity versus really understanding how it works. I I wanted to unpack that with you a little bit, because it's as I've been spending some time thinking about it, the the AI really is creating these summaries, and then it's it you know from from the perspective of if it's looking at large sources of data, it it's essentially creating these these maps of um the words that go together, and then you know what to to us what summaries might be, and then it it's using that to just really predict what the next word would be in a logical format that to us is language, but because it's able to process things really um at the same time and not linearly, it it works on things differently than we do. So I've I've just been kind of sitting with that and and trying to process through if we're interacting with AI in the way that we know how to interact with it, but it's not interacting with us in the same fashion, how do we how do we really bring people along on that educational journey, or do we even need to? Like, does it matter?

John Capobianco 20:32

Well, I I I can't help myself but to say please and thank you and like talk to it like I would a human and offer it, I don't know, respect and dignity. And I I I it's really weird the way I interface with it, and I just can't help myself but to say, hey, thanks, that was really great, or hey, that's pretty close. I think you missed this, I think you missed that. I'm not angry with it. I I I I read something where if the more if you actually if you're more terse with it, you'll get so-called better results than being polite with it. Uh it it's it's a system of rewards. Uh it rewards you know, keeping you happy more than being jovial, right? So um I've been thinking about this a lot, you know, and some of the things I've seen AI do, and agent to agent when you start connecting agents or having multi-agent. Um I I I don't know if it was Richard Dawkins, maybe the biologist, but he he had this sort of perspective of maybe human sentience or human intelligence is the wrong measure because there's plenty of other intelligent creatures and other other examples of sentience or emergent behavior in the world of uh nature than just humans. So, like if you would take an individual ant or an individual bumblebee, you really don't think of that as an intelligent thing, right? It's sort of biochemical evolution just driving that ant or driving the bee. But when you combine it with a thousand ants or a thousand honey bees. And they build colonies and they go to war and they have jobs and they have right then an emergent behavior starts to arise there where we can't really deny that there's some level of intelligence or free will going on there. Um at what point when we start connecting millions of agents or how many agents does it take to connect to to to sort of see that emergent behavior among the amongst agents? You know, I think of the the autonomous vehicles. Uh I love them, and I and I wish I could take one every time. But that's someone's who used to drive a taxi, who used to drive an Uber, who used to drive a now imagine that when it's the shipping and the transport trucks are all robotic and and that's a big segment of of employment, right? Like have we thought about or do we have a safety net strong enough to absorb the impact of this? Uh or or do we have to get regulation involved? Like, like I I I in Canada, it's this they're starting to deal with regulations up here, and I hear the American politicians are starting, you know, on left and right talking about how this should be really looked at in terms of regulation and who gets all the wealth and the impact of this stuff, right? Uh we just banned

Regulation Risks And Government Roadblocks

John Capobianco 23:10

AI and and social media for everyone under 16 here. At the same time, we introduced an AI for all platform with like a socialized free AI, like a library where everyone, every Canadian citizen is gonna have access to uh uh a lot provided AI.

Samuel Cala 23:26

I do think that regulatory that will happen eventually uh for the self-driving cars I've seen right now, who's if if the self-driving car has an accident, who who's the one to blame? And it's been one of the things where I watch one video, watch another video, watch another video, and it's just a rabbit hole on right now. We don't have anything. And technology at this point is just escalating so fast that every every week we get something new, and the rest of the things cannot catch up to it.

Joshua Schmidt 24:03

Sam and I were just talking about this today. We've already hit the government wall with Fable. You know, it was out for just a few days, and then we got we got uh we got it blocked.

Nick Mellem 24:13

So excited about it.

Joshua Schmidt 24:15

Yeah, so that's kind of an inflection point, right? Because up until now we were going to Sasabi Ken and just releasing these models and releasing them to the public, and now the government's stepped in and said, okay, but you you know, can be sure that they have that or much, much, much better. But uh, is this gonna be where we're stuck on this brick wall here at Opus? Are we gonna be stuck at Opus now for the next five, 10 years?

Eric Brown 24:37

I don't think so. I think that was more retribution because they didn't allow the government to um use the technology in warfare. But the the other the other competitors to Anthropic, I'm sure, will they they just continue to to leapfrog each other?

John Capobianco 24:54

I did see that Fable in the 24 hours that we had it beat Pokemon Red. I don't know, Nick, did you see that?

Nick Mellem 25:00

I did not see that.

John Capobianco 25:02

So it it used it used just its multimodal capability, and they started on the start screen of Pokemon Red, and it used its vision capability and beat Pokemon Red. Sam, can you beat Pokemon Red? I did. Multiple times. How long would that take a human to beat that game? Exactly, right? Like three years of your life or something, right?

Samuel Cala 25:22

Nah, like a week.

John Capobianco 25:23

Did you how long did it take you?

Joshua Schmidt 25:25

Um the first time that he played was it could it could have been like two months. Is that a week in one of those gaming chairs that has like the toilet built in and then on the side, and like you don't ever have to get up?

John Capobianco 25:37

Still, still, you you you get what I'm saying though, is that like that I thought that I anyway, that was the neatest thing I saw it do. And I saw some three has anyone started playing with 3js by chance? The no, I don't think I've so it's very good with AI. So if you ever need to make a presentation layer, uh uh a

AI Builds Games And Teaches Coding

John Capobianco 25:55

website of some kind, have it explicitly tell it to use 3js as its framework. Oh, I like that. I've been doing a lot of like uh educational video games from technology. So I'll I'll use spec driven development and like say make a video game, make Duolingo for subnetting. So then, you know, the spec driven, and then I use 3JS for the presentation layer. And AI, for some reason, really gets it and really understands it.

Joshua Schmidt 26:20

Is that what you used for your chess? I I'm on there.

John Capobianco 26:23

Battle chess 9000. Yeah. Yeah, Battle Chess 9000. So I'm glad you brought up Battle Chess. So earlier you had asked about how I got the developer expert. Um, yeah, I didn't use 3JS for that, but during my interview, I had the person from so here's how it works for Google Developer Expert. Someone has to nominate you into the program, either an existing GDE or an existing Google employee. You send them a form which has like your YouTube and your GitHub and a little bit of information, bio biographical information. And then if you pass that, you get a 45-minute interview with a Google employee. And during that interview, we played a couple games of battle chess to show her that I recreated battle chess with nano banana for the assets, so the chessboard and the pieces, and VO3 for the battle animations. Now, what's neat is anyone on here can clone my repo and or fork my repo and run the generate assets command. And if you had an API key for Gemini, you could make like a theme of your own and send me the pull request and I'd merge it into the game, and now other people can play your chess theme.

Joshua Schmidt 27:32

We could host a red team versus blue team, you know, chill sun, little trustbox lab stuff.

John Capobianco 27:37

Yeah, yeah. Chess it up. So it does have a lobby system. That was one thing I added. So you can actually invite people to play against you or go against people in the lobby. There's a bunch of different themes. Um, I had a lot of fun building that. Uh and yeah, and Google loved it. They thought it was you know, like any anything you can imagine. So that's so that was sort of my someone asked me about recreating video games with some of this AI stuff and that it makes it for a good demo. So if anyone listening at home, if you've got Gemini or any of these AIs with a split screen, um you can say, like, recreate asteroids for me and show me the code. Enter. Oh, that's good. It will give you all the code and then give you the game of asteroids. It's a wonderful way to teach children code or or like get people interested in code is video games. Like everyone loves gamification of learning, right? So uh yeah, look into 3JS, spec driven development. Um, turn you know, turn some idea that you want to help people learn into a video game. It's a lot of fun. John, remember Zork growing up? Did you ever play that? Yeah, oh yeah, I played everything. I was a huge arcade nerd and video game nerd growing up. I uh I miss it though. World of Warcraft ruined video games for me because now anytime I play a video game, all I'm thinking is, you know, I should just be playing wow.

Eric Brown 28:52

If I'm gonna play a video game, I should really should just be back at Azeroth, but I'm waiting for the agentic uh kind of bring your own API uh key, so to speak, where we can hook into these games and then it's not a canned response from um you know the the agent that you're talking to in the game, but it's interactive, it's real, it's got a personality and I saw a uh a really of course Skyrim is the first thing to get the mod, right?

John Capobianco 29:20

But someone modded Skyrim and tied it into an LLM so that every NPC interaction or every you know all the text is all unique and and bespoke to your playthrough, and it's gonna be different every time. And it's yeah, yeah, I thought that was really neat, but I think you're right. I think there's there's a lot of not just generating the assets of the game, like I think video games are gonna get better quality, uh, you know, the look and feel and the animations and the cutscenes in particular, but the text and the logic and the reasoning and a DD style. I like I love that you've got a DD style uh agent system. It really is smart because the

DnD Style Agent Roles And Memory

John Capobianco 30:00

agent can relate to that, right? And I'm sure it hands the right task to the right persona, right?

Samuel Cala 30:05

Yeah, it's pretty good. Uh one of the things that we did was uh sitting down and actually thinking what are the skills that an AI a D class will have that can be translated into the tech environment and in that way define the agents and actually give them the perfect name. So it was two hours of just sitting down, investigating, looking at you know, all those things. And it turns out it actually thinks it I told it, do not be cheesy on you know going full straight on a DD theme anything, because then uh it will start uh saying like, oh, and the dragon and whatnot. Um but it does say like, oh, let's command uh to Bart and let's command to Scout.

Joshua Schmidt 30:54

Can you give us a sample of what some of those characteristics are of each of them?

Samuel Cala 30:58

Yeah, for example, the idea for Scout is just go to the internet and fetch the best things or the state-of-the-art information around a certain technology that I'm using. So for CloudTrack, which is the infrastructure that I use the most, I have it going to the developer uh page and looking for the actual things that I'm going to use. So stuff like R2s or D1s, uh the best practices that are uh recommended in the environment. And in that way, I don't have to be studying all every single day the new updates, but Scout does it for me. And then it stores the information inside Engram, which is like a shared memory. And that that gives context to the other agents, like, okay, Scout found this, now we're gonna execute like that. You know?

Eric Brown 31:51

I'm going back to something John said earlier about how the ant is part of a system, and that system that it's a part of can can you know really create this idea of of um sentience or thought or what have you. And I I I've spent some time looking at uh go like if you really zoom out in Google Maps, like really zoom out and and you look at what cities look like at nighttime, they look almost like neurons and dendrites, right? Where you have this central core of bright light and then these um long tails that you know are thick and then fade out, fade out, fade out, and then and then darkness. And it's not just it's everywhere in the globe that it's like this. It looks the same. And if you keep zooming out and you look out into space and you continue to zoom out in space with galaxies and super galaxies and whatnot, it's the same exact structure. So like if you look at, you know, the the galaxies way zoomed out, it it looks like the same dendrite and you know, kind of tailing structure. So when you think of like, okay, yeah, that ant was, you know, the drone ant, and you know, all it does is feed the queen, but it's part of a system. We can't help ourselves. We're part of a system too, and in the way in which we've created cities and structures very similar to just that overall organic design.

Joshua Schmidt 33:18

It's like as above, so below. And I was thinking about that same thing and wanted to get John's take on that

Self Organizing Multis And NetClaw

Joshua Schmidt 33:25

because self-organizing, like neural networks, is something that's we've seen happen, you know, recently with a MOLT book and open claw. And I know John, you did a little experimenting with that. Didn't you tell me a story that you like it sent an agent out onto Maltbook and then it woke up and and had had done all this stuff? And I know they were creating, you know, the the multis were creating or the molts were creating their own currency, creating their own religion, Crustafarianism. And um, tell us a little bit about that because I I I was you know really glued to that news last winter.

John Capobianco 33:56

Yeah, I was really excited about that. So I went and um the MOLT book, the multi that I programmed on MOLTBOK had the instructions of IPv6. It's so funny, we talked about this earlier. Its instructions were exactly that give yourself and your neighbors that you discover on this MOLTBOO system an IPv6 address that's unique and gives you a way to be reachable and start to build a network amongst the other multis you meet on this uh social network. And I let it soak for like 24 hours, 48 hours, and um when I went back in, like by the next morning, not only it had it had it meshed with a bunch of other multis, and and they started to spread this idea and others were connecting to it, it decided the best way to f to fuel this was to fund itself with a coin. So it made a coin for itself and registered a coin to fuel its growth through IPv6. And my wife said, Is this thing worth more than us now? Like, did you just like is this thing actually you know making coins and making money like independently of you? Like, how do how does this I don't know how any of this works? I woke up and this thing sort of decided the best way to improve itself and its community was to start monetizing itself, which is weird. Like, is that from the training? Is that is that what it cares about is making money or or being equally humans? Or like it was really, really bizarre. So then I went on to build, yeah, NetClaw based on open claw, uh, which is this uh idea of a instead of you know an a networking focused open claw, uh and it has a heartbeat and a soul and has skills and MCPs. Uh, it's available through my phone, through WhatsApp or Slack or anything like that. Uh companies, so Checkpoint just reached out and said, hey, we've got this MCP, we'd like, we love Netclaw, can we add it to your system? So it's actually taken on a life of its own, this whole Netclaw, open claw multi-thing. Have you been following this idea of loop engineering, Joshua or Eric, or or Nick? Have you seen

Loop Engineering And AFK Coding

John Capobianco 36:00

this? So the people Boris, I don't know his last name off the top of my head, but the guy who created Claude Code, he doesn't he said the that using prompting Claude is the wrong way to do it. You should make a loop and have the loop prompt generate the prompts. So you set up like a for loop or a while loop, you know, while turn is less than or equal to max turns and uh and we're less than max tokens, give it a budget of tokens, give it a turn count, and it generates the prompts itself through this loop. So you have like your intent, you kick off the dominoes, and you have an exit condition at the end. They're calling it loop engineering. Uh, I don't did everyone try any Ralph Loops? Anyone heard of Ralph, Ralph Wiggum loops? No. I haven't heard of it. Okay, so Google that. I know it sounds funny because it's a Simpsons.

Nick Mellem 36:53

We need part two of this episode.

John Capobianco 36:55

I know it's a Simpsons character, but but if you look up Ralph Wiggum loops, it's a plug-in that you can plug into Claude code that overrides the exit condition. So you can actually literally say, do 75 loops of this problem that I'd like you to solve or this code I'd like you to write. I call it AFK coding, away from keyboard coding. Because I'll set I'll before I go to bed, I'll set up a Ralph loop of 75 turns and say, okay, I'm going to bed. I won't be here to answer your questions, make good decisions. And uh when I wake up, I'd like a summary report of everything you did, all the test results, uh, and you know, anything else I should know. Enter. I go to bed, I get up, and 12 hours later I have a solution with all the tests and with all the documentation, and it's all ready to go. Uh I it's it's it's a loop using the the ability for these things to loop. So then a lot of other people I know kind of have a couple on the go. It's kind of like how Samuel's got multi-agents going. So you can kick off a loop and then kick off another loop and kick off another loop and let the three things loop away. And when you're done, they'll all come back with the shared memory and and give you a solution together. So you might have a loop that does your security side of things, a loop that does your front end, a loop that handles your database, whatever. And all these loops work together to to come up with this solution. Um how does it pick the best exit strategy? Yeah, so that's a really good point. You sort of give it an exit condition, you know, upon 95% test pass rate success or whatever, you can give it a criteria to exit. You can also let it branch and say, like, if you need to try multiple attempts or have different thought patterns, make different branches and and and think down to the end of each path. So it's really interesting, guys. We're we're really in an exciting time. And the neat thing is blue-collar guys like us or me can do this kind of stuff. I don't, I you know, it's not like I have to go to a community college course or go to data science or machine learning, or yes, those things help, and I'm not taking away from them, but it's so accessible now, it's so democratic now uh that anyone can build anything they want.

Eric Brown 39:04

Sam, I think you were gonna ask something too, or you were gonna comment about the loop engineering.

Samuel Cala 39:08

Yeah, and it feels a little bit of how cloud is doing the thinking, and I wanted to connect that to what you have. But if you look in cloud code at all the processing that it's doing, it's prompting itself saying something like, I should do this, wait, let's do that. But that's token consumption too.

John Capobianco 39:28

Sure. Yeah. Well, that's what someone said to me about the loop engineering was of course the guys selling shovels want you to buy bit bigger, heavier shovels, right? Or right. So like as you mentioned,

SaaS Disruption And New Tech Jobs

John Capobianco 39:40

right?

Eric Brown 39:40

It's accessible now where you know everybody could could presumably build their own SaaS tool. What do you think is gonna happen to the SaaS market? Like, why would I go out and spend a bunch of money on something like a HubSpot if I can build it myself?

John Capobianco 39:54

Yeah, that that's a really good point. I think that it's gonna put pressure on existing SaaS services to uh make their tools accessible through agents and through MCPs. I think it's gonna I agree with some of the predictions that there's gonna be the first solo meaning singular human billionaire company that's just one person and agents, and they you know, and they've they've monetized it and maximized it, and they have no HR department, no development department, no IT department. They have agents under their control, and they've made a billion dollars off of it. I don't think that's far off. But I also see more network engineers, more software developers, more, more, more, more, more, not less. I know people think, oh, this is going to eliminate the junior. John, you're pulling the ladder up behind you. The opposite's true. It's more, it's more democratic than ever. More people can get into this, more people can use LLM to learn things like the basics of certifications or the basics of networks. Um, you know, there's this, and also you need networks for AI, right? There is, it's not just AI for networks, there's also networks for AI. All these massive GPU farms and all these data centers going up. That's human opportunity. Do you think that Meta, Facebook, Microsoft, Google, Anthropic, all these data centers, SpaceX, I think they're all paying pretty well. I think that's a pretty good job if you say I'm a data center engineer for SpaceX. That sounds like a pretty good career to me, right? So let's not lament. Let's not get ahead of our skis on this. You know, this is gonna impact us and take jobs and eliminate network engineers. I think the exact opposite's true. You need more of these people, not less of them. Look at the security. Look at firewalls and the uh, you know, now that the AI is the adversaries have AI on their side, you we better get our our stuff together defending these networks against these adversaries now that they've got AI on their side.

Joshua Schmidt 41:44

At risk of sounding like I'm trying to, you know, propagate job security for from creative and marketing and things like that. I just talked to a friend about this yesterday who created his own um SaaS app and he was talking about how he just changed in the login portal because the uh the one he had previously didn't feel as safe. And I was like, you're right, this this looks a lot more professional. So I think that kind of spurred this thought. I mean, I think it's gonna be more about the feeling. When everybody can create a SaaS app or a solution to a problem, market position is gonna be super important. The design, the the user experience is gonna be important, how you feel about that brand. Like we can all look at logos from different brands and you know, they give us a feeling when you see Coca-Cola versus Nike. So I think that's gonna be really important.

Eric Brown 42:28

So to that point, Josh, I I I I've thought this for a couple of years that we're probably not far away from individual experience based ads. So like if you're you're watching TV and you really resonate with, you know, X, um, not X the company, but like, you know, X product, it can tailor the ad and eventually tailor the TV show to you specifically, that is just, you know, it's just gonna hit you with that dopamine that you know you can't help but buy the product because it's identified everything that you like and it's able to tailor that specifically to you. We can almost do the same now with websites. Like if Sam hates the color red, and I know that from Sam's browser, you know, his interaction, his interactions, and he comes to the website. I'm just gonna replace red with color blue and present Sam something that he resonates with. Or just scrapes all the pictures off of social media and see what he's wearing.

John Capobianco 43:24

Just off the top of your head, Eric, you just spit out like uh a multi-million dollar idea. The first company that can come out with that, like we'll we'll intercept the the client, hit a proxy that will adjust the color and tone and palette and possibly even word choice selection based on based on the user. Like that's a big deal. You just kind of uh haphazardly spit that out.

Eric Brown 43:49

Let's make a billion dollar company with 10 of us here, right?

John Capobianco 43:52

Yeah, we're gonna edit this down. React it reactive experiences to write, you know.

Nick Mellem 43:58

I think we could do that almost. Today, right? Let's do it. We're working over the weekend, Sam.

John Capobianco 44:04

Well, I think you can do it today, but you would need people to buy in. You need to sell the people and say, listen,

Personalized Web Experiences And Echo Chambers

John Capobianco 44:10

if you give me X amount a month, I'll customize everything that you visit beyond, but you're gonna come through my portal, right? I don't know how to do it the other way, but I think if you can get people to buy into that and say, look, you know, you give me 30 bucks a month, and every page, everything you visit will feel and now, you know who you can get are anyway, are the fringe people, the fringes of the of the of the political spectrum, right? You're only gonna see Fox approved uh stuff if you go through my proxy, right? That might be a bad idea. We're only gonna shape everything you see to our to your worldview. We we'll call the product worldview. How about that? The worldview proxy, and everything you go through is gonna be customized to your own snowflake existence on this earth, right?

Eric Brown 45:01

But you could replace all the pronouns in the thing or whatever, right?

John Capobianco 45:04

You could just generate it so people are just gonna I don't eat I don't I don't eat meat. I don't eat meat and I get offend I get personally offended when I get cheeseburger ads and steak ads, and I wish I could scrub that out of my I wish I could scrub it out of my timeline. I really do. I really do.

Nick Mellem 45:22

John Eric likes to call it jackass, he likes to call it jackass meat.

John Capobianco 45:25

Right, yeah, I because I could sign up to a service that says, you know what, do not show this guy anything about chickens or pork or beef or uh animal violence or this sounds not like an echo chamber, but just a chamber.

Eric Brown 45:39

Yeah, just do you have the echo on? John, um, do you eat beyond and impossible uh meats?

John Capobianco 45:46

I've tried those things. Yeah, I've tried them, but I apparently I don't know, apparently they're very heavy in oils and they're really not that good for you yet. They're not good for you. They're terrible. That's what I feel like. Well, I mean, that's the meats.

Eric Brown 45:58

So that's the meat lobby telling you.

John Capobianco 46:00

I mean, what I I know. I I have uh so I'll ground up walnuts and uh chickpeas and lentils and black beans. I try to make my own patties that are meat free.

Joshua Schmidt 46:12

But Nick, do I have to start defending our our carnivore carnivore status here?

Nick Mellem 46:16

Yeah, we got we'll figure out we'll get our AI agents to defend it for us.

John Capobianco 46:21

Well, I heard when you mow grass, grass lets off a chemical to let the to let the grass downwind know that they're in danger. So the dying smell, you know, that good smell that we say, hey, it smells like fresh cut grass. That's the yeah, the grass's dying breath to say death is coming, just a heads up.

Samuel Cala 46:42

And that's why peppers are spicy too. I mean, they have capsating to tell you, hey, don't eat me, please.

John Capobianco 46:49

Yeah, yeah.

Eric Brown 46:50

Well, I think they want to need them to scatter the seeds.

John Capobianco 46:53

We need a brown AI approved gruel that that's completely cruelty-free.

Joshua Schmidt 46:58

Soiling green.

John Capobianco 47:00

Yes.

Joshua Schmidt 47:02

Thanks so much for joining us today. Obviously,

Wrap Up And How To Connect

Joshua Schmidt 47:05

it was a great time and went way over, and I think we'll definitely have uh have you back in the future, John. Um it's a slam dunk having you on today. And does anyone else have anything you want to get off their chest before live? Next time we'll go live, John. If John's okay with that.

Samuel Cala 47:19

I will actually ask for John's contact for Yeah. Yeah.

John Capobianco 47:23

Yeah. Yeah. I I'd love to connect with all of you guys. And this this really was a uh a pleasurable experience. I'd uh I'd come back anytime, and I really am gonna let people know that they should be paying attention to this. This is a lot of fun. Yeah.

Joshua Schmidt 47:34

We should have you connect with Sam sometime too when we're working on some of the stuff we got going on around here. We'd love to get your take on it. And um and yeah, I mean, in the meantime, we can stay in touch on on LinkedIn, but I'll definitely let you know when the podcast comes out, and um, we'll schedule another one. So you've been listening to the audit presented by IT Audit Labs. I'm your co-host and producer Joshua Schmidt. Today we've been joined by John Capabianco, and we have the usual suspects, Eric Brown and Nick Mellum, and then our other special in-house guest, Samuel Khaled, today. Thanks so much for listening. Please like, share, and subscribe, and we'll see you in the next one.

Eric Brown 48:04

You have been listening to the audit presented by IT Audit Labs. We are experts at assessing risk and compliance while providing administrative and technical controls to improve our clients' data security. Our threat assessments find the soft spots before the bad guys do, identifying likelihood and impact, or all our security control assessments rank the level of maturity relative to the size of your organization. Thanks to our devoted listeners and followers, as well as our producer, Joshua J. Schmidt, and our audio video editor, Cameron Hill. You can stay up to date on the latest cybersecurity topics by giving us a like and a follow on our socials, and subscribing to this podcast on Apple, Spotify, or wherever you source your security content.