This is the art of network engineering. Where technology meets the human side of IT. Whether you're scaling networks, solving problems, or shaping your career, we've got the insights, stories, and tips to keep you ahead in the ever-evolving world of networking. 00:15 Welcome to the Art of Network Engineering podcast. My name is Andy Lapteff and in this episode, we are joined by an old friend of the show, John Capobianco. Hi, John. you doing? Andy. Hey, Michael. Good to be here again. It's been a long time. Really happy to see that the show is thriving after all this time and all those years. Yeah. Thank you for having me. This is like your third or fourth time back on. Thank you so much for coming on. We have some really exciting. 00:38 uh fun, provocative stuff to talk about in this episode. Before we get into that, my bearded friend, do want to give a plug to your employer? And then I think you guys do have an MCP server that you created. Do want to give a quick shout out to the work you guys are doing? Yeah. So I'm with Itential now as the head of AI and DevRel and their MCP server, their technology is one of the first in the industry. And a big reason why I joined, Peter Spurgata's work and a few other people, Joxon and Ankit and William Collins, really remarkable work. And it's secure. It has OAuth 2. 01:07 It has our back. has everything you would want in an MCP. And another reason why I joined is that this, that there have a new product coming out in the summer flow AI, where we actually get to build and deploy agentic infrastructure agents. So it's everything we've been talking about wrapped up into a job. I've really enjoyed it. The culture there is quite amazing, you know, and it does work very well with my previous employer selector. So there's no hard feelings there. It's just a different opportunity for me. Let's go over to Mike Bushong. Hi, Mike. How are you guys doing today? I'm super excited to talk to you. 01:36 Are you really excited? I am. He said John is like larger than life. And for those of you who haven't met him, you think that he's larger than life because he says things with a lot of energy. And then when you meet him, you realize that he's actually not just figuratively larger than life. He's literally larger than life. 01:51 And he gives the best hugs. Yeah. Everybody's five foot seven on zoom, right? So, uh, yeah, it's a, it's a bit of a shock when people actually meet me in person to see that I, uh, I live up to this ambience and this energy, but I know that there is probably a certain segment going, this is all going to be about AI and this, you know, all the cool stuff and why I'm falling behind. And I want to just maybe set the stage and say that this is intended to be a positive experience for the listener and for the panel here tonight with. 02:20 optimism and hope and a utopian outcome. And we're going to say some provocative things. left my Autocon speech on a pretty provocative note, but it really is to inspire you to maybe look at things a little bit different, to consider the tools available to you. And that's all that artificial intelligence is. It's still a tool, right? It's a cool tool, but it's still a tool, right? 02:42 So John, thank you for doing my job for me. You are the man. That's going to be our topic of this episode. Really the working title in my brain is ABCs of AI. And really I have adopted the use of an LLM in my daily life. I'm a chat GPT guy. I signed up a year or two ago. I use it all the time. And for the most part, it helps me a lot. It's good at certain things I needed to be good at. It seems like there's AI everywhere. There's AI toilets. There's AI pen. it for me. 03:09 I don't want to call it cynicism. don't know what the right word is, but in my network engineering, you know, groups of people, anytime anybody says AI something, there's like this proverbial I roll on like, oh my God, okay, now you have the AI sticker too. So what I would like to come away with in this episode is if you're a network engineer and you are surveying the landscape of modern tooling, what you would need to stay employed, stay valuable organization, stay skilled up. 03:36 right in your career, where does AI fit in all that? And really, how do you separate the hype from the reality? And John, we were talking right before the show, you and my friend, Juan Lightfoot, you two have always seemed further down the road than the rest of us. You can seem to like see around the corners and you're always at the place that we're all like, oh no, that looks really important, we need to catch up. So here we are again, trying to catch up with John and all the things you've learned in AI in the past year or two. So I thought where we'd start is, 04:05 Something I heard you say at Autocon4, it's provocative, but I think it'll get folks' attention and get them engaged. The Network Automation Forum, did their, I guess, annual survey. And the 2025 survey showed that I think it's something around 70-ish percent of networks aren't automated at all or automated meaningfully. And John had said at the end of his amazing talk, which was, think, from CLI to GPT, you said the industry won't upskill 70 % of folks when the 04:34 AI LLM machine learning. Again, I don't even know what words to use, so you're going to educate us. But when aogenic AI gets so good that it can do a lot of our networking jobs for us or better than us, 70 % of those people who aren't automating at all will be displaced. And how did you put it? Like we're all going to be out of jobs. So you better skill up now. Right? That's was the net. This is nuts. 05:03 It's going to be positive. So John, I know what I was trying to get at. So first of all, I was shocked at the number. I was optimistic again. Okay. This is going to be the year like the year of the Linux desktop, 2025 is the year of automation. Finally, there's been a breakthrough and the numbers were disappointing again. 05:22 And it's 10 years going on 12, 13 years of some of these automation tools, RESTconf and NETconf and Ansible and Python. And I was let down by that. And then I sort of thought, but what does that mean in the age of AI? There is 30 % that are doing that. And I used to think that the 70 % was an opportunity to teach people new skills and to upskill them and make it 50-50 and then make it 70-30 the other way. But I think the 30 % 05:49 are clearly using it. I'm glad to hear that you're using AI and have been for a couple of years now, Andy, at least even in learning how to interface with it and how to prompt it and what is a good use case and what's not. I really think that the availability of things like agent development kits, ADKs, relatively low barrier to entry, things like fast MCP to make an MCP server out of a tool, it's democratized a lot, right? 06:16 In terms of what, how good these models are, the latest Claude 4.6, Opus model, Codex 5.3, Gemini 3, these models are very, very good now as foundational models to the point that they are equivalent of an average CCNA, CCNP. If you have the right approach, a CCIE or a JNCIE or an Azure expert, or you name your expertise level, right? They're very capable of doing these things. 06:46 So with the right platforms and the right guardrails and the right approach, and we'll talk about things like rag and things like MCP tonight, by the time this podcast is aired, there's probably going to be like a million new agents on the internet, right? Like they're, they're proliferating at an extremely fast rate. And from a corporate point of view, right? What I've heard is that it's not, AI is not going to displace the people. It's people who are using AI are going to displace the people that are not. 07:13 Right. And I make a, maybe an assumption there. Maybe it's a bold assumption or incorrect assumption that if you're not doing network automation, you're not using AI, but that to me was my trajectory. was a natural evolution of my phase of life doing automation. Right. I'm curious when you look at like automation and sort of the take rates, I guess, do you attribute the low take rate primarily to, we've been waiting for the right tool. been waiting for the right piece of technology, or do you attribute it to. 07:40 You know, there's a people problem that we have, whether that's training or incentives or, or fears or confidence or whatever, right? There's like, guess, how do you view the tool versus, or I guess tech versus people, which is the bigger inhibitor to progress? think it's probably a 30, 30, 30 split between the solutions offered by the industry itself, all vendors included. I'm not going to pick on one individual vendor, but something that you have to know net comp and. 08:08 templating and a domain specific language. And we've tried to unify with things like Yang models, but there's three different flavors of Yang models and there's a lot of friction around those using them. So I think some of it is, the actual tooling that has been available for the last decade. And I think that there's a cultural aspect to this at the individual contributor level, a sort of a mindset of, listen, I went to school to learn networking. I've been getting networking certifications and now you're asking me to become a programmer. 08:36 and a proficient programmer and put production on the line with code that I'm not comfortable writing. Right? Like there's years to become comfortable doing automation and production that that takes some time to learn things like get and VS code and all that stuff around the tools. And then I think there's just the, general enterprise approach, whether or not they're investing and saying, let's do an automate first approach, right? Next year, we're to put a line in the sand and everything after that date must be automated or. 09:04 Every new project that comes up, we're going to start automating with that new project, right? I think those three have led to the 30 % number, right? There's also questions about scale. Are some networks too big to automate? Are some networks too small to automate? How do you automate or move things to the cloud or CICD pipelines in an air-gapped environment? We're a blue chip company. We care about stability, not innovation. 09:27 There's a lot of competing factors here, right? think there's two things that I want to kind of pick up there. And before we move to the tech side, I have a belief, I guess, that the networks, the way they're constructed today, tend to be quite fragile in a lot of environments. And so even if you could go fast, it's like you wouldn't because you're so afraid of change that you have these draconian change controls and you have blackout windows that extend from before Thanksgiving, well into February. And so those 09:56 to me don't scream, we need a tool. Those scream to me, we're afraid of doing anything. And if you can't take away the fear of doing something, then the tool is almost irrelevant. And so I think part of it's that. And there's probably, you know, roots in architecture and in change management. And then I actually think the tools have a huge role to play in making sure that people don't make mistakes. Like there's like a big, a big piece of that. The other piece I think is, interesting. And this is where, I think, I think what you're going to do in the, in the next, however many minutes we talk, I think you're going to take away this next one for me. 10:26 I think the industry dialogue is always around, you know, automate or die, right? Like you're asking a workforce to implement a thing that they fundamentally believe is aimed at removing their jobs. And I think when the narrative is like that, it's like, you can't tell people that, you know, Hey, please build your demise because you won't see the kind of the broad adoption of it, widespread adoption of it. 10:48 because people are afraid. And I think the way you've positioned it already, you so far in this call or in this podcast, like it's opportunity based. And this, you opened up by saying, this is like a positive, a story of triumph. This is not a story of misery and defeat. And I think that's the piece. Like if I can wave a wand and change anything in our industry or on automation or on operations broadly, it would be that. Like it wouldn't be a tool. would be like, give people the fundamental understanding that like, look, what AI is doing on top of bringing in a bunch of other, you know. 11:15 capabilities is this driving like just a monstrous amount. 11:18 traffic and spend and whatever. And so we can't grow the teams linearly with the CapEx. And so at some point, like the AI to me is like the, it's sort of the thing that allows us to even cope with it all. I would be more afraid of not AI than I would of AI because like my job is going to be terrible. If you're telling me I got to manage, you know, four times the capacity and devices and whatever, and I'm going to have the same head count. Like I think there's hope in that, but I don't know, I guess I don't know how you view it. then, you know, clearly you're going to have to figure 11:48 How do you get those people enabled so they can do something? I agree with you on both points. I think the first one is a little bit of a self-fulfilling prophecy in that networks are fragile because we don't change them ever. So they're fragile because we don't change them ever. Right? Like I think there's a little bit of a self-fulfilling prophecy there. If we never touch the network and we never progress or change our methodologies or the tools we use, Like, know, putty is still around things like that. 12:15 I think there are better ways now. Here's what I'd like to see network engineers do is pick the top five or 10 things that they do every day. And you used to think of a way to orchestrate them in a workflow with automation in some way. think now the opportunity is to try to make a personal assistant, an agent, a co-pilot, a coworker. There's many ways to look at it that is under your management. Right? Like I believe that this AI agent issue is an HR problem. It's not a technology problem. It's how do we incorporate these agents into a 12:44 Orchard properly and with very closed guardrails and specific defined permissions. You know, the company I work for now has 200 people. If every one of us could make five agents that worked for us, we've just added that many more people that work for the company, even though they're not physical beings. Like you said, we just can't keep throwing people at these things and the time it takes to become an expert on these things versus. 13:10 You know, some of the approaches of making that knowledge available. Like I think it can really augment the senior network engineer who really knows their stuff as much as it can a junior engineer who is learning their stuff. think there's a spectrum here and AI is as good at or better if it's going to multiply your skills by a hundred. If you've got zero skills in that thing, that's still zero, right? But if I've got 70 or 80 % on that scale of expertise, that 100 times force multiplier is really powerful. 13:40 I think I see what my role is going to be in this episode. So if 70 % of our networking experts won't automate and I think your statement was just that, you know, AI can augment the experts, then if you're an expert and you're not automating, I think it's a fair connection that you made earlier, which is you're probably not using AI either. Do those experts go away? And then where does our 14:02 tribal knowledge go in networks. It's a circular argument I have in my own head. I'm optimistic that I think they will adopt this. I don't think that they will hold out. I think that sort of patience has run out around this topic and now there's no more excuse, right? Like whose patience if they're not going to automate who's going to make them do AI, I guess is my I'm being provocative here. I'm not arguing with No, you're right. You're right. I'm not trying to argue. It's a good point. I don't want to see 14:30 the displacement. I'd rather see them embrace it and augment themselves with AI to write those scripts that they were previously held back on. There's sort of a jump to the front of the line effect here. Maybe it's a good thing I didn't focus on learning Python over the last 15 years, because now an AI can write better Python than if I had spent the time learning Python. If you're looking at it through the lens of, I'd like to automate something on my network and I've never had the skills to write the scripts or the Python to do that. 14:58 Now you have this expert that can give you tested code using your source of truth and using your rag and using MCPs. So for example, there's literally an RFC MCP server for the request for comments, right? So if you plug that into your agent and say, I'm writing OSPF code, use the standard to help me guide my config. It will not hallucinate. It will not get it wrong. It literally has access to the OSPF. 15:26 IEEE RFCs. And that's just one example of an MCP that I found very useful doing network automation code because it's truly referencing the standard, right? There's many, many uses around this for network engineers. One, to like your, you know, when you had Eric on talking about learning the code through AI, right? But it really should be a co-pilot, a co-worker of you, right? So there's certain tools that I think if you haven't been already using, 15:53 VS code and Git and some of the foundational network automation tools, the water is deeper for you to start using AI just from scratch, right? And I understand that some people play with it for 15 minutes, it gets something wrong or it hallucinates and they're turned off by that and they have a bad experience and they sort of dismiss it. There's many deeper layers when you start to use it programmatically through the REST API or connecting it to agents or connecting it to model context protocol. I keep coming back to MCP because 16:23 I feel like that was a really big necessary step and tool in the ecosystem of AI to kind of standardize around a protocol. Right. So it's, it's done for AI, what SMTP did for email or what HTTP did for the web. Right. Like I really think it's that big of a, of a protocol because it allows you to provide that context to the model through JSON RPC. just a transport mechanism. Right. 16:49 You can wrap anything from a REST API call to a Python script to a database call. Anything you could think of can basically be presented now as a tool that the agent can call upon. So when I wrote a PyTS MCP, so that way I could just say to the agent, you test the health of these interfaces? And because one of the tools the MCP has is a show command tool, it can just call that show command tool and execute whatever arbitrary show commands. 17:18 the agent needs to run. Like that truly is autonomous Andy. Like I could set that agent to monitor the health of a network, to configure a network, to make documentation, to write tests. Are there guardrails to constrain that from doing things you don't want it to do? I'm guessing there are. So there are like markdown format files where you kind of outline. 17:38 Here's a good example, right? Please do not update the enable secret. Do not lock me out of this device. Do not add an ACL that's going to prevent me from reconnecting to the device. So yes, you, and that's where the human expertise comes in, right? To Michael's point and to your point about senior network engineers, who would know that other than a senior network engineer to put those guardrails in to an AI agent, right? It's still very much augmenting a human expert that would know the nuances of some of the protocols. 18:03 VTY lines, enable secrets, default routes, management interfaces. So yes, there is, there is guardrails that you put in place in the agent, way you build the agent itself. think that the, the role of the network engineer. So if you're mop and SOP based or you're not thinking about the workflow, you're not thinking about the architecture, you're not making decisions, you're executing what's written for you and you're a set of hands. It's not hard to see those people replaced by scripts because you're literally executing what somebody else says. 18:32 documented for you to execute. You're like the deployment arm, if you will. I think when you know workflow, and I don't mean like basic workflow, like, you know, provision this, ping that, but like when you understand how these things come together, then, then you're in a position that you can use your co-pilot or your agent to do something because you have this unique insight into what has to be done. Every tool in like an operational ecosystem is like, is a, is a window to a workflow, right? It provides some information or it lets you. 18:58 you know, execute something or do whatever. And the question is, never how does a tool work? It's like, how do you work across tooling boundaries? How do you get like multi-dimensional things happening? And I think the network engineer has a, has a, like an operational architect role to play there. And there's a lot of things that you can do. And then as AI gets better at translating, you know, some of the workflows into 19:19 code or applications or whatever, then you have, your toolbox is bigger. You've got more things that you can go do, but I don't think it obviates the need for someone who actually understands like, what are you actually trying to do? So I do think there's value. Like I think the people who really know workflow, the people who really know operations, the people who can look at something and discern what's going on, I think there's value for them. And then I think it's a different story if they've relegated themselves to read prompt, type prompt. agree with you and that can elevate them. 19:48 they do becomes more strategic, it becomes more focused on the governance and all that, the, you know, not the sexy stuff, but I think that you're right. And those people should, that's what I'm trying to appeal to network engineers listening to this of all stripes. 20:02 and of all experience is to try to start building your own agents that can do some of this work for you. The other thing is with the reasoning and tool calling capability, it's quite remarkable what an agent can do. You could have a ticket in service now and through the MCP, right, pull that ticket information in. And then like I mentioned the PyTS or there's Ansible MCPs, there's all kinds of config management MCPs out there. Have the agent pull the ticket, read the ticket, address the situation on the network. 20:29 be another MCP, send a Slack, send an email, right? Do all the things with a high degree of quality. that the information you get in an email that's AI generated. Yes, it's obviously AI generated, but it would take a human hours just to craft the email with the details you get, right? 20:47 Like what this is about rapid and moving faster and resolving issues quicker. And that's sort of Andy, you had some questions earlier about examples of out of the human, out of the loop. Here's, think the path that people should take is start with human in the loop, have the AI prompt you this, as this the config you want to deploy and use your reasoning and your, your expertise to say, yes, proceed. Right. I wouldn't even start with configuration management. I would start with read only activities. Show logging is a great one. It's just pages and pages of text. 21:16 And humans are not very good at parsing logs. They just, we've never have been because of the nature of the logs. So why not take that dump of logs and either just put it right in the context window, put it into a vector store for rag. So you can look up, you know, logs in a vector store, make an MCP that you can use to do show logging on a certain platform, right? Config management, your intent versus what's actually configured. It's very good at text deltas. 21:43 Right? So config management, not just pushing configure generating config, but compliance, right? Things like that. There's tons of read only stuff you can do testing. What I, what I think is neat about the PITS thing is that if I'm asking about BGP tests, will dynamically know what commands to run and what tests to generate. Like we've gotten so far that the AI agent is now actually generating the tests and looking at the results of its tests and giving you the evaluation. It's not even running. 22:11 handwritten deterministic tests anymore on the fly, it's creating those tests and running them, right? And this is what I think network engineers should be building. This is why I'm so passionate about this is those are the tools I think that people should be building now to augment themselves. Imagine interface help tests, and you just say, run the agent or in Slack or all my interface is healthy on the core, enter. So that's sort of what I mean by vibe ops. And it's funny, someone said to me today about vibe coding, in the near future, it's just going to be called coding. 22:41 Right? Like everyone does it. Everyone does it that way now. Right? So to distinguish it as vibe coding, I think it's almost past say, I think it's just called coding now. Right? Yeah, there's, I want to take us in a direction and I could spend days talking about any one of these things, vibe coding included, because I've tried and failed multiple times, but a couple of things I want to say, and then I'd love to steer us into some definitions and some terms that at the end of this conversation, if people know three or four 23:09 terms and definitions from you and maybe how they can apply them. You know, at the end of this conversation, how can you get into this world and do some of the things that you're talking about? So real quick, you were talking earlier, and it made me think of this TM forum, like levels of autonomous networks, which I had never heard before until last year with with a co worker who brought it up, you know, it can sound very tin foil hat of me to be like, Oh, no, what are we gonna take all our jobs and as gonna take all our jobs, but I do think that if our networks can get 23:37 to the point of reliability where it's just like the power flowing or the water turning on. That's probably a good thing for everyone. So I think a lot of the things that we are talking about that my instinct says, oh no, it's gonna make me lose my job. If you skill up, you can stay in that environment and then you can help those networks run more reliably like power and water. So just one point to try to reorient my biases. The other thing you said I liked was the time to learn. 24:03 It takes years and an enormous amount of effort to understand networks enough to be able to unleash a human on them, right? And then years to learn all the automation tooling and encoding and all that stuff. And now probably years to ingest all of this stuff. So I have a lot of newbies reach out to me, hey, I heard this thing and I've been following you and I'd love to get in and what should I learn? And my advice five years ago, it was different than two years ago. And now today it's like, well, 24:29 learn all the routing and switching, learn all the programming and now learn all this AI. Like I would go running if I were them, but we just keep piling on more and more onto network engineers and it's tricky, but here we are. The last thing I'll say is something about fragility. And I really want to get into some AI terminology and technologies. You had said before, you know, don't touch anything. And it like my experience in production was we had standards, which our industry would call validated designs and very smart people wrote them when we were supposed to build all of our data center networks to follow the standards. Yeah. 24:58 And then there were exceptions driven by business units who had this very important thing that had to be deployed yesterday. And the timeline has been bastardized, blah, blah. So we make exceptions and put in a temporary solution to build this thing that the BU needs right away. And then that thing stays there forever. And if you are a big enough company and you do that enough times, now you have one-offs and snowflakes all over the planet. Good luck trying to make 25:25 those networks reliable, stable and have expected outcomes when you touch them because everything's a snowflake. So again, not that there's anything to say there other than I think that's what drives a lot of the fragility is not following your standards or your validated designs. What I really want to get into LLMs, MCP, Agenic AI and RAG. I just have a list of like terms. So for a network engineer who wants to get up to speed on all things John and Duann and these AI experts talk about, like where do we start? Do we pick an LLM and just start? 25:53 And then how do we apply these things to networking? like, let's bring it together here, tooling, and then how do you do network stuff with it? First of all, I would not do any shadow AI, right? So if you're on your corporate asset, follow your corporation's guidelines and reach out and find out, am I allowed to use co-pilot in VS code? For example, do we have an enterprise subscription with be it Anthropic, be it Google, be it Azure, be it whoever, and can I get API keys? Right. 26:19 through my company. On your own at home, I would suggest you start with Olamma, LM Studio, or Microsoft Foundry. Now those are three open source private ways for you to download models onto your device and start doing inference. This is for like zero, day zero, I've not done any of this, how do I get started? Before you invest in a public hyperscaler model, which is a subscription fee, 26:46 which is better quality typically, and you can do cooler things with them. But if you just want to get started locally, by the time this podcast is over, you could literally have pulled down a language model and be chatting with Olama. Olama has a REST API locally. So if you want to start writing code and writing agents, you can use Olama or LM Studio or Foundry. That's local that runs on your GPU, right? On your Mac. Olama will actually run in a Raspberry Pi. It's very, very small, lightweight stuff. But is it dumb? 27:14 It hasn't been trained. doesn't have like, how does it know anything? is that in the model? That's just the framework. So that's, that's, that's where you're going to use to run models that you then download from the cloud. And in terms of models, and I'm, I'm maybe behind way behind on some of this, but you know, there's models from Google called Gemma three. That's a Google open source free model. can download five four from Microsoft PHI four Mistral Lama three point three. There's various models that you can download. 27:43 question just to try locally. Yeah. Why would you start locally and not just do like a free chat GPT as an example? What's the benefit of doing something locally? I think that network engineers generally feel more comfortable doing things offline and there's a security and there's a privacy concern. any, so that I, I, why I recommend starting there is because if that is what's holding people back and they think they have to turn over their data or pay the $20 a month, you have a broad 28:10 range of people watching and listening to this, right? Some people just don't have 20 bucks a month to invest. Yep. But from there, I would go AI shopping. It depends on what you want to do. If you want to do code, if you want to do image generation or video generation, there's a lot of different uses and some models are better than others, but you can't go wrong with chat, GBT, Gemini, anthropic meta even write the X model or the grok. mean, so 28:40 Yes, I would find one and invest the $20 a month. And from there, you're going to get an API key. And that is really the key to all of this is starting to do programmatic stuff. Yes, you can use it to help you polish your resume or send that email. But when you want to plug it into something like co-pilot in VS code or, you know, anti-gravity or cursor, there's many different IDEs that are AI capable. And then you can start doing the cool stuff, right? 29:08 Help me write an Ansible playbook that can we pause one second. So the 70 again, I'm, this is my role here. I guess the 70 % of folks who don't have VS code running don't understand programming. So I don't know if we just hit the point in the conversation where if you haven't learned anything in coding and you haven't installed VS code and you don't know how to do a co-pilot add on in VS code, is this the point of the conversation for the listener of like, shit. Okay. I just hit the perimeter of. 29:36 What I know, what I can do and now I'm going to tune out because I don't know VS code and you get what I'm saying. Like how do we pull them along? So do you have to know programming to do what you're saying? I think so. You're making a good point. think maybe I jumped too fast too far in that you're going to spend some time in cloud desktop or chat GPT or Gemini in the browser. Just 29:56 getting your feet wet with prompt engineering. Why is the sky blue? Maybe copy and pasting things out of your router or out of your logs into the context window. Help me understand these logs and you just dump in some stuff there. Like I want to bring that 70 % along with. I know you do not have to be a programmer to start doing this stuff. Well, let's say for those people, like what was that agent you talked about where you could do something and it would go against the RFC? Like, like to me who isn't a programmer, right? Like that sounded compelling when you said that earlier. Oh wow. Okay. 30:26 Okay, so the first thing you're going to want to do is to try to connect that model that has a knowledge cutoff date and think of it as a closed book exam. You're going to want to give that access to your resources or external resources, right? So that's where the model context protocol would come in. So in Cloud desktop, there's a simple settings developer tools and a button that says edit, right? Edit the JSON file. 30:52 And here's where it does get a little tricky for people. If you've never worked with JavaScript object notation, if you don't know how to read JSON, if you've never worked with it, unfortunately- I also don't know what Cloud Desktop is. Like this is where- Okay. Now I'm like, uh-oh, now my butt's puckering. So Cloud Desktop is Anthropics version of Chat GPT. Okay. So it uses a different set of models. It has a different kind of look and feel to it or Gemini from Google. So those are sort of the three chat interfaces. 31:21 The good news is that MCP being a protocol is agnostic. So that RFC MCP that I mentioned, you can connect it to any of the tools I've mentioned, chat, GPT, Claude, Gemini, VS code, co-pilot. So you plug in the JSON and now suddenly your Claude desktop has direct access to the RFCs and you can start asking questions about OSPF or BGP or whatever. As an example. Okay. Time out. Let me ask a question. So. 31:51 So John, what's the hello world? So most people use, you know, chat GPT as an extension of Google, right? They just put in search stuff and then they get things out and then they, and then they do, they add questions to it. So they get maybe a richer answer, but they don't even use that to do like real work. And this is partially why there's like a huge disconnect between what AI is capable of because people had that experience like a year ago and they were using it like kind of ineffectively and they were using models that have, that are many generations past. 32:20 So now there's like this world of capability, but you got to use it in a different way in your mind. Like what's the hello world to get people away from. I'm going to use any of these sites as a, effectively a different search bar. I here's a pretty universal hello world specifically for model context protocol. So if you've gotten to the point in this discussion where you have cloud desktop or a co-pilot and you want to connect your first MCP, I would use net box or not robot. Both are equally good. They both have. 32:49 And MCP, that's the important part. And you can use their sandbox environment. So do you do not have to use production? I would recommend staying away from production. Actually use the demo net box site or the demo not robot site, issue yourself an API key. Okay. Now that might be the first time you've done that, but you're going to need that key and then see if you can connect their MCP to your, I'm going to say agent because all these things we've been talking about our agents, they're agentic. So I plugged that net box MCP into my agent. 33:19 Let's say miraculously get it working on my first attempt. Now I can start to ask how many circuits I have in Atlanta. What IP addresses are free in this certain subnet? You can literally start talking to your source of truth. That is that Michael would be my hello world because it's sort of a universally used tool, a source of truth. Infra hub. There's other ones. I don't want to leave anyone out of this discussion, by the way. I'm just using two as an example, but. 33:46 Find something like that you can sink your teeth into. And it really isn't as hard as a JavaScript and this and that. It really isn't as hard as I'm making it out in the cloud desktop. You're going to paste in the link to the MCP on GitHub. And you're going to say, help me understand how to install this, right? You're going to use AI to help augment AI. I wouldn't do any of this without AI. Here's the MCP server. I want to install, help me understand how to install it. Walk me through this because I'm a network engineer, not a programmer. 34:14 I've never used JSON before, right? Help, help dumb this down so that I can install the Netbox MCP in the Claude. And then you can start adding records from Claude, right? I'm going to provision a new site tomorrow. Here's the details, enter. And now suddenly those details are populated in Netbox, right? It really is that simple and that smooth. had a visceral reaction to what you said, because in my job in production, someone would ask me how many circuits do we have in the Phoenix data center or how many such and such routers are we running somewhere? And now I'm. 34:42 grabbing spreadsheets, trying to find them in a SharePoint, logging into the jump box, spending hours jumping around to make sure they're updated. If I could have a conversation with my source of truth in natural language, like when you said that, I got tingly. I'm like, holy shit. Well, what if, but what if you applied, what if you built that agent, Andy, because you are the expert in that stuff and you just said people go to the Andy Jr. Just, just ask. Right. They're never going to get rid of me. 35:12 Right. mean, wow, the value I just created. Hey guys, you can talk in natural language to our source of truth. You're welcome. Right. I'm going on vacation. So the other thing is, and you know what? I, Michael, I wanted to pick up on this earlier, actually. It almost slipped my mind, but I've always been rewarded for automating myself out of, out of a job. Like when I went to 35:31 to my leadership of parliament and explained to them that we were going to do this, you know, 50 router upgrade and it was all going to just be a playbook. They didn't say, well, we don't need your help anymore, John. Thanks for doing all that. Right. They actually said, what could you, what else can you automate? Are there other things that you're thinking about doing? It actually elevated me quite a bit. I was never fearful for automating myself out of a job. feel the same way about the agents, Andy, if it's Andy's agent and you call it Andy's agent. 35:58 Right. You've established real value and, and, and it made everyone's lives easier. Just talk to the bot, right? You got a question about an IP or a subnet or a site or a circuit. Well, we've just wrapped it in an AI agent that can fulfill those needs. Right. make a good point. There's so much mess to clean up and automate. I don't think they get rid of us. I just interviewed the St. Jude network architects and they were telling me about, uh, they automated their device replacements, you know, the switch replacements, and they went from six months. 36:26 to a stack of switches to like four weeks with an automation thing, but they're all there, right? They're not all gone because they're automating all the stuff. I'm glad that you shared that. Like they weren't like, oh great, John, go away now. We don't need you. Cause there's just so much, think, to clean up and economies that we could. 36:43 create with automations. When Dave Ward and I were doing a bunch of work in kind of the early network programmability days, that's what we called it before it became SDN. So we were terrible at marketing, but we were like very, very early. I told Dave and we actually worked it into a keynote that he gave it at some crazy, like some huge events. I told Dave though, the network program, same idea, right? It was network programmability going to drive people's jobs away. I'm like, no, it's like a lottery ticket. So I think there is a continuum of time. 37:10 So if you get there early, it's a lottery ticket. It's opportunity. If you're the last one to the party, then you know what? The keg is going to be empty and there's nothing for you. And so I think there's people need to feel a little bit of urgency because I think the faster this train goes, the harder it is to jump on while it's moving. And so you're better off going in now. I like the hello world. Like I like the idea of doing something that's relevant to what you're doing day to day. It's super bounded in terms of what can go wrong. 37:39 because it takes away lot of the fear. It's super constrained. Like you're only talking about, in this case, two elements that you have to kind of pull together. You're going to use natural language. So there's not like some heavy syntactical component to it. Like those elements, like that's what makes it a great hello world. And the people who can do that, like the way they should view it is, you know, yes, you're picking up a new skill or you're dabbling in a new tech, but you're purchasing a lottery ticket. And the thing that's cool is you can purchase more of those lottery tickets. Like the more you kind of double down. 38:08 And so you have a chance to participate. Like there will be wealth created during this, this boom and not everybody will participate because not everybody gets off the couch, but you can, and it doesn't have to super scary. I think it's very much a bit from a zero to a one that will flip in people's minds. Maybe they just haven't had that personal epiphany, that personal inflection point. It happened to me quite early. 38:30 right connecting network automation to chat, GBT 3.5 that bit slipped. And I, I've drifted away from my networking roots and spoke as most of my energy. Now I'm trying to keep up with artificial intelligence over about three years. So it's not insurmountable. It's not, I know it's, it's intimidating. 38:49 Thinking, I'm going to go from networking to artificial intelligence. Don't I have to be a computer scientist to do that? Don't I have to know machine learning to do that? Think of it in terms of the tactical using the tool, right? You're not going to be training models or fine tuning models or building big GPU farms. Some of us will, but most of us are going to be using this to augment ourselves in our day-to-day jobs and trying to build agents to reduce the friction. Now, once you have that net box agent, 39:14 The next thing you're going to do is say to yourself, hang on, there's a service now MCP. I could snap that into this agent and now start making tickets in service now through natural language, referencing the net box data. our prompts become more compounded. We're actually invoking multiple tools in one paragraph of instructions to the agent, right? I want you to then send an email. 39:36 I then want, I'd like a Slack. I'd like you to make a PowerPoint out of it. There's unlimited possibilities, right? We're only bounded by our imagination at this point. It really is about how creative you can think about the availability of this. They're not tools in, it may be like a toolbox, but more like a palette of paint. There's all these different colors. There's all these different technologies and you can dabble in here and do a little bit of rag, do a little bit of MCP, build an agent. And now suddenly you've painted this wonderful picture. 40:06 Hey, john, MCP is the protocol that allows, I guess, intelligent systems to talk to each other. Is that so it's client server. So the server might be like, uh Google Maps just released us cloud hosted MCP for Google Maps. So if anyone wants to do cool things with Google Maps, you used to have to read the API spec, you used to have to use a tool like postman or write coded Python with a JSON body and a post. 40:33 All of that has been replaced now with just natural language. So I could plug in my, into my client. And we've been talking about clients tonight, cloud desktop AI agents, co-pilot. Well, I plug the MCP server into that client. And now I can say, where's the nearest Arby's right in my co-pilot. And because it's got the access to the map MCP, it will give me an answer based on the. 40:56 The last time I checked, was over 17,000 MCP servers on a clearing house. they free and people make them to like netbox made one, Google maps made one. Yeah. Now for the Google maps one, need a Google API, but people want their tools integrated into their tools because that's what drives you use. So what you're going to see is like a proliferation of connectivity options. And then what they're trying to do is if you can reduce the barrier to 41:23 doing something cool with access to your tool, then consumption goes up. like vibe coding, like to me, it's like you used to have like an R and D team that sat between you and an idea or your idea and execution. Now what you have is a set of agents or a set of tools that sit between you. your ideas have never been more accessible than they are today. Right. And every day they'll be more accessible than they were the day before. That's very insightful. So Andy, here's an example, Andy. There's a WordPress MCP server. Okay. 41:53 And my, my automate your network.ca blog has been hosted on WordPress forever. So what I did was I added that MCP server to my Claude code. And at the end of one of my more successful vibe coding sessions, I said to it, can you please do a blog writeup of everything we've accomplished here together and posted on my WordPress site. And I pressed enter. And 35 seconds later, there was a multi-page blog that was referencing all the stuff we had done during my vibe coding session. 42:22 I'm sorry, who did you give that instruction to a WordPress MCP server? I plugged in the WordPress MCP server into my Claude code where I was doing my vibe coding. And then at the end, I just said, use the WordPress MCP server to publish a blog about the code we've written here today. Andy, let's be clear. What he's basically signing you up for is that you're going to go through and you're going to do a hello world and we're going to record it. We'll edit it down so that any 42:50 mistakes along the way. Like you want to, you know, it's almost like speeding it up, but that's like, if you want to show people that hello world is possible, you should show people that hello world is possible. I'd be happy to join and guide you through that. And we could actually use the net box dev site and, uh, make an agent, whatever. think we need to show, and I know you're doing this, John, this isn't, this isn't a gap of content. 43:14 That's so impactful to any network engineer that you wouldn't even need to know coding. if, think the hell the world is that Netbox. Let's do it. Let's central it up guys. on. honestly, need VS code and copilot and you get free, a free, certain number of free calls to copilot. So that's free to get started. The Netbox MCP, JSON code and an API key. And because it's ephemeral and because it's not production, we can share the key. can, you know, do all this in public. 43:42 And, uh, and we'll do it from, we'll do it from a couple of places. We'll do it on Claude, desktop. Yeah. It's really cool. We're going to document it on John's blog because we're at the end, you're going to connect it into, uh, his WordPress site and the WordPress MCP here. Yeah. Mike, you Mike, can you talk to my boss and clear the deck? I got stuff to do. I think this is fun. I'm not going to type, but I'm going to be on the, I want to, I want to be watching. 44:09 because I'm not even a network engineer. So I look at this and this is, this is cool. And let's do it. Let's just say yes. Yes. Yes. John, I got questions. Yeah. And actually, you know what, you know what a good MCP for you Andy would be is the YouTube MCP because all these transcripts of all your videos could be accessible through natural language. Thank you for saying that. So my vibe coding attempt and failure and hour of my life, I'll never get back before I had Eric of the devil on my show. 44:37 I wanted to pull her nine transcripts from her YouTube Learn to Code with AI videos. So this is one way I use LLMs. Pull down her transcripts, give me a quick summary, tie it in and help me build an episode on, my God, look what you've built. This is beautiful. An hour with ChatGPT, it's spitting code at me, me using it, me getting an error, and then it gaslighting me and telling me, oh no, this is great. It's okay. We're almost there. An hour of my life, And this is something that, 45:07 Dwan said, if you don't understand the code it's giving you, you're helpless. And that was my experience. So give me code, put it in, get error. Give me code, put it in, give error. Now I'm a human in the loop that is probably in the way. And I don't understand, oh, the API thing is all, oh, the YouTube agent is this. like, so I do not enjoy that experience because I don't know what I'm doing. I don't understand it. And Chad GPT loves to just tell me. 45:34 We're almost there. We're so close. Now it'll definitely work. And it's so full of crap. It doesn't. Now I know that Andy, you're going to have a guy and what we're going to do is we're going to do it live and we're going to invite your subscribers. They can join. They, everyone can just, they will all be the audience, not watching and judging you. We'll be cheering you on and experiencing it with you. that MCP server works better than vibe coding, I'm in is I guess the point of that rant because MCP servers. 46:03 It will actually augment your vibe coding. part of your vibe coding, it'll actually make it a lot easier to vibe code. Yeah. It really will. Bad vibes. When I vibe code, it's all bad vibes, man. So Andy, we're at 55 minutes in. I think you have like where you want to go on this. John, for the people who are still questioning, like what would you leave them with? 46:31 We've made an impassioned plea. We've said the things that you want to do, right? The pain you want to alleviate. The thing you always wish was better is never been more accessible than it is today. You've made a great, a bunch of suggestions around where might you get started. You've given people some very specific tips on different tools that they might look at first. And you've actually given a bit of a sequence and then we've committed at least loosely to then walking through that hello world. 47:01 How far do you think this goes? Not like over like a five year window, because I mean, who knows five years, right? But let's say by the end of the year, even for somebody who's starting today, what might they be capable if they put in a reasonable effort given the normal constraints of day to day life? One thing to just be optimistic about is all this MCP stuff we've been talking about is it's literally one year old, right? So how far behind can you be? 47:27 You're not that far behind. It's, it's one year old and most iterations of it didn't come out until nine, eight, nine months ago. So the train is still in sight to Michael's point, right? But don't let it drift too much further ahead without trying model context protocol, because I think as network engineers, we should be excited by a new protocol. We live and die by BGP and OSPF and RIP and 47:49 HTTP and all these different protocols. Think of it that way. Think of it as another protocol in the digital era that's going to help make your life a lot easier. Another, a couple of other points. I sort of think we're in, we're, kind of beyond slop now. I think these models are so good and the code they're creating. So 20,000 Claude agents work together to create an operating system. So that's how smart they are right now is that 20,000 of them together can actually create Linux. 48:18 in just a couple of days. So ask yourself, is there going to be more or less of this tomorrow? Right. And I know that there's fears of a bubble, you know, I don't know. I think maybe, maybe like a dot com bubble that leaves behind all the fiber and all the telecoms and all the infrastructure, maybe less like the housing bubble, but I don't know. Even, even that I hear less, less about, less talk about the bubble. Just get onto the train, just find that one thing. 48:45 that will help you if, if you had an agent be able to do it for you. Right. And Andy and I are going to get you set up with how to start with MCPs, but you've heard me talk a lot about different tools today. One is better than none. Right. Don't be hung up on, you know, analysis, paralysis on the tools you should start with. And the other thing is you really do have to sort of keep your ear to the ground. It's moving very fast. Every 49:09 day or two, there's new groundbreaking announcements, new models, new capabilities, model context protocols are being announced. And maybe even look at something like Cisco. Cisco has added MCP to their automation exam. So they get a CCNP in automation. Now you have to know about model context protocol. I've been looking at the news coming out of Cisco live uh in Amsterdam and every session has to do with AI MCP agents. think that 49:35 the tide has shifted and to my earliest point about the 70 and 30%. Like I said, Andy, think that generally speaking, enterprises are starting to run out of patience with the network, right? It can, it can no longer continue to be the bottleneck or the source of pain or the reason why things go down, right? Another BGP issue, another DNS issue, right? I think that patience is going to be lost. And to your point of they want running water, they want taps that turn on and off. 50:03 Right. They're not too interested in the nuances or the minutiaes of network engineering. They want a utility that works. Right. And, and Michael, you said it earlier, our systems are so complicated now. What better time in human history for AI to come along to help us understand and untangle some of this mess. Right. I love it, John. Thanks for coming on. I learned a lot. Where can people find you, John? If they have more questions, if they want to reach out and learn more about. 50:27 Sure. So I'm on available still on X Twitter, first name, underscore last name, uh, LinkedIn. I'm easy to find. My YouTube's pretty easy to find. I've been a little slow on YouTube videos lately. Uh, but that's just consequence of a new job and some, you know, new responsibilities, but I hope to get making more videos soon. And there's a new vibe ops forum that we started in uh a Slack channel. So if you're interested in, in sharing your work in a safe. 50:51 you know, open space that's embracing AI and there's about 500 people in there already. You can find out how to get into there on my socials. Mike, where can people find you? Cause the new thing is people come to me to get to you and I'm finding out of that loop. You need an agent for that Andy. I am the agent. Yeah. I gave up on the socials. I'm mostly on LinkedIn these days. I couldn't handle the, uh, the socials. I'll just leave it at that. The constant doom scrolling kind of got to me. 51:19 So reach out to LinkedIn, anyone who sends me a note, I respond to basically everything. And so if people need help with stuff, you let me know. I can't always help, but I can usually direct to other people who are smarter than I. Thanks so much for coming on guys. This is a great conversation. For all things Art of Net End, you can check out our link tree at linktree.org slash art of network engineering, all kinds of fun resources and things there, including our discord server called It's All About the Journey. Thousands of folks in there lifting each other up when they win and patting each other on the back when they lose, which usually means failing an exam. 51:49 Home Lab channel, just all kinds of wonderful things happening in our community. So you don't have to trudge this road alone. If you don't have a community, you can check us out there. As always, thanks so much for listening and watching, and we'll catch you next time on the Art of Network Engineering podcast. Hey, folks, if you like what you heard today, please subscribe to our podcast and your favorite pod catcher. You can find us on socials at Art of NetEng, and you can visit linktree forward slash Art of NetEng for links to all of our content. 52:15 including the A1 Merch Store and our virtual community on Discord called It's All About the Journey. You can see our pretty faces on our YouTube channel named the Art of Network Engineering. That's youtube.com forward slash art of net edge. Thanks for listening