(So you can skip ahead, if you want.)

00:00 Welcome and Introduction
07:59 Why Context is King for AI
14:18 Agent Experience and Safety
18:12 Itential Platform Introduction
20:27 Human-in-the-Loop Journey
27:29 Connecting Netbox MCP Server
32:30 John Builds an Agent with FlowAI
38:06 Bidirectional Sync Demo
42:25 Getting Started with AI
50:32 Local LLMs and Model Selection

Kris Beevers • 00:04

Hi folks, you’ve got John and Kris here. A lot of folks are piling in right now, so welcome. We’re going to give it a few minutes for everybody to join before we kick off, and I’m going to go ahead and Get some slideware up. What would be really helpful is if anyone who, yeah, is on can just chime in on the chat and say hi, make sure we can, you can hear us. Great. And Zach, yes, this will be recorded. Hey folks, we got a good crowd today.

Kris Beevers • 00:53

I’m really excited. Um, okay, I’m going to share my screen.

John Capobianco • 00:58

And yeah, this is, uh, very open. Um, we did put some planning into this, but it is really for the community. We, uh, Kris and I have both a lot of really valuable things to share today, and we, we got together. Kris saw some of my work. I’ve seen some of his advancements in Netbox. And we thought, why don’t we bring some of this to the community and have a nice, open, inclusive, safe space to discuss where artificial intelligence intersects with both his work at Netbox and my work at iTential. Today’s going to be called Closing the Context Gap Between AI Agents and Network Infrastructure.

John Capobianco • 01:38

It really is about the context. Let’s pay close attention to the context. I’m going to be putting this into practical terms today. I’m going to show you some actual Flow AI agents using the Netbox Labs technology. Together, agenticly. So Kris, yeah, thanks for having me. What are your thoughts about today’s session and what people are going to see?

Kris Beevers • 02:06

Yeah, thanks, John. Yeah, I mean, I think you set the stage really well right there. We’ve got already an incredible crowd. I see we’re pushing 200 people here live at the moment and climbing. So excited. What I’m going to do, just for everybody’s benefit, is take us through a couple of just webinar basics and then, you know, talk a little bit about how we see this world and this theme forming up from the Netbox Labs point of view.

Kris Beevers • 02:38

But I think the main event is what Jon just described. Like, we, I think, want to get hands-on, have Jon show you some of the agents that he’s built, have some, as you said, Jon, some like participation and discussion for which we’ll mostly use the chat and just have a nice, fun, organic session here today. And so, you know, with that, I feel like we’ve got enough, definitely got enough critical mass here at the moment. Really good group of folks listening in, and maybe I’ll just get us kicked off. And John, feel free to interrupt or chime in with anything, but I’m just going to get us introduced and, you know, pretty quickly get it over to you. So folks, again, we can see you on the chat. So if there’s something you have a question about or you want to comment on, don’t hesitate.

Kris Beevers • 03:27

I expect a robust discussion in the chat today, just given the topic and how many folks are here. Obviously, all of us are thinking about this theme. of agentic AI and infrastructure operations. Now we’re going to dig into it in a ton of depth. Before we do that, here’s a look at how we’re going to run it today. For anyone who’s not familiar with Netbox, Netbox Labs, we’ll touch on that real quick. I think John will take that kind of my potential intro a little bit later.

Kris Beevers • 03:57

And then I, again, I do want to talk just a bit about how we see, you know, the themes around agentic AI and infrastructure forming up from our vantage point at Netbox Labs. What we’re working on in this area and then give it to John for, I would say, the main event where, you know, he’s going to take us through Flow AI and some of his agents. And if we have time, maybe a bit of a live demo. And I think we should have time for some discussion and Q&A as well. So that’s how we’ll run it today. Just a couple webinar housekeeping things. I already mentioned a couple of these things.

Kris Beevers • 04:34

Use the chat. I see folks are using it already. Welcome. So, you know, we, we have the opportunity for some live discussion. This session is recorded. We will send it out to everybody afterward. That’ll take probably a couple days is my guess.

Kris Beevers • 04:52

You have any problems, there’s an email address there, marketing@epoxylabs.com. During the session, team is monitoring and we can try and help you debug, but hopefully this is all pretty straightforward. We can’t hear you if you’re speaking, so all interaction via the chat, please. Okay, and I’m just going to go real quick on a couple introductory things. I think almost everybody here probably knows Netbox, but if you don’t, Netbox is the world’s most popular network and infrastructure source of truth or system of record. If you don’t know Netbox Labs, we’re the commercial stewards of Netbox, and we deliver the Netbox Labs platform, which is a whole host of features and products and operationalization around Netbox, who’ve been around a while now, and obviously Netbox is super, super mature. I’m not going to spend a ton of time on that or this, but just so you have a sense of us, you know, Netbox Labs, our platform spans the entire lifecycle of infrastructure network management and operations from as early as what is in my infrastructure, you know, documentation and planning and design, all the way through, is it working the way I intend it to?

Kris Beevers • 06:09

Can I get more business value out of this data? And then obviously driving change, and we’re going to spend some time on like the AI side of automation today. And so, you know, pretty wide array of value in the Netbox Labs product portfolio if you’re not familiar, and always of course we’re happy to talk with folks about this stuff. Anytime. Why do people work with Netbox Labs? All of that value, and of course, you know, support, scale, additional product, and so on. So just wanted to take a couple minutes for folks to be oriented to Netbox Labs because I think almost everybody’s already familiar with Netbox.

Kris Beevers • 06:53

Now I want to take a few minutes in a little more depth to pivot toward Agentic AI. And I’m sure there’s a lot of folks on this webinar that have been on other webinars over the last 12 or 18 months that I’ve done on this topic. We have been very focused on AI in infrastructure, AI in network, over the last 18 months and understanding Netbox’s place in this world as the world is evolving really rapidly. And one theme that has held true from sort of the outset of this modern surge in LLMs and agents is that context is king. You know, it’s really hard to effectively leverage AI in infrastructure tasks or really any kind of tasks if you don’t feed the AI effective context.

Kris Beevers • 07:59

And it turns out that’s what Netbox is with respect to, you know, Networks and Infrastructure. Netbox is a, you know, structured dataset about your footprint, you know, that touches, depending on the nature of your infrastructure, every component, all the relationships between those components, metadata about those components, how they’re configured, change history, all of this knowledge that forms— we sometimes will use the term a knowledge graph or a semantic map or a context map or something like that, that grounds agents in the context they need to be able to reason and act on your infrastructure. Without that kind of context, agents run wild. They, you know, can run far afield. And I think John will show us why this is important and how it really fits in in practice in a few minutes. But that’s the way to think about Netbox and the Netbox data model in an agentic AI world.

Kris Beevers • 08:57

It is the context that enables these agents to reason and act. If you’re not familiar with our work in AI at Netbox Labs, we got a lot going on in this world. And so just a few things I want to touch on quickly, you know, with this sort of grounding thesis that the context in Netbox is critical to agentic AI. We have a three-prong strategy, and the first is make it really easy for agents to work with that context. So John’s going to talk about, you know, how he’s integrated Attentio’s Flow AI agents with the Netbox MCP server. We’ve had an MCP server for a long time at Netbox Labs and always maintaining and investing in that. And why have we built that?

Kris Beevers • 09:49

Because we want to make it really easy for agents to work with the data in Netbox, and this is a prime interface. We also have libraries of skills for agents to work with Netbox. And in general, this term here you see, AX, we think a lot about agent experience when it comes to working with the data in Netbox and the broader Netbox Labs platform. So one pillar of our strategy is just make it very easy for agents and AI to work with your infrastructure data because we believe connecting that context is super, super important. A couple other pillars of our strategy to touch on real quick. interactive agents. So I’ll touch on this a little more depth in just a second, but we’ve invested in building interactive agents to work with your network and infrastructure data on a kind of minute-to-minute basis and accelerate workflows.

Kris Beevers • 10:42

Netbox Copilot is available to everybody today. Again, I’ll look at it a little more depth in just a second. And then a third pillar for us is autonomous agents, and I think this is roughly where John is going to take us today with Flow AI. This is a huge theme for us at Netbox Labs. How do we enable autonomous agentic workflows that are safe and that can, you know, accelerate the work of teams, take away grunt work that no one really wants to be doing, and generally enable us all to uplevel and focus on more strategic priorities. So that’s our rough strategy. Just a quick side note or detour, if you haven’t tried Netbox Copilot, or you’re not aware of it, it’s worth just a brief call out.

Kris Beevers • 11:32

Netbox Copilot is GA as of a month or so ago. We first announced Netbox Copilot, I don’t know, probably last summer, I think. It went through an extensive preview period. Let’s play that again. Preview period, a ton of refinement, very mature at the moment. It’s an interactive AI agent that is in Netbox, and you can use it even if you’re using Netbox Community Edition, not Cloud or Enterprise. There are some capabilities of Netbox Copilot available for you today.

Kris Beevers • 12:04

Just go to your preferences, turn on Copilot, that’s it. It’s like a 2-step process to enable it. It’s worth being aware in Netbox Cloud, Netbox Enterprise, and actually that’s what this thing is showing, Netbox Copilot has a bunch of more advanced capabilities. Most notably, you, you can make changes to your Netbox data safely through Netbox Copilot review and approval process, but Netbox Copilot can accelerate actual change in your infrastructure data. And then of course all kinds of data governance, data sovereignty type options for enterprises. So Netbox Copilot, pretty mature, worth checking out if you haven’t already. Just a couple more things and then I’m going to get it to John.

Kris Beevers • 12:53

One thing that we’ve come to learn at Netbox Labs and really believe in that is guiding our investment strategy is that all of your tooling that you’re using to manage your network, your infrastructure, as you move forward is going to need to focus not just on, you know, your experience as a human user, but also on the ability to connect agents and enable agents to work with your infrastructure tooling, your infrastructure data. So another important element of our strategy at Netbox Labs, and something that I would encourage everyone to be thinking about as you’re examining your tooling ecosystem and, you know, preparing, I guess I would say, for an agentic future that is happening real fast now is what is the agent experience like with respect to your tooling, your operational environment and ecosystem? How do you make it possible for agents to address the workflows that you want them to take on, on a day-to-day basis? And this looks like You know, leaning toward and choosing tools and systems that are API-first or that have first-class MCP ecosystems generally are programmatically addressable. That’s really important. And then, you know, I think another big theme for us

Kris Beevers • 14:18

is that the user experience for network and infrastructure operators is going to change in the next few years. What are we going to care about? Probably not as much pointing and clicking around tables in Netbox and, you know, editing individual devices. Probably we’re going to ask the agents to take care of that more tactical work. And what do we need to uplevel toward? Planning and design workflows, more strategic review of the infrastructure, visualizations, analytics. Those kinds of interactions are going to become much more important for us as operators over the next few years.

Kris Beevers • 14:52

And so this is how we’re thinking about evolving the Netbox platform now in preparation, I guess, for like an almost immediate agentic future. And then the last thing I just want to touch on, and I think this is maybe a segue to, John, some of the grounding you want to share with everybody is, you know, AI is not for free, and we really need to think about safety in introducing AI and agentic workflows into our systems, especially as we’re operating critical infrastructure environments. And this really comes in a few forms. And those of us who have used AI for, for example, like software development use cases have really felt this. Coding agents, as an example, are much, much more effective when you give them rules when you give them the ability to self-validate. And these same principles, they’re going to map to agentic infrastructure operations as well.

Kris Beevers • 15:56

And so what do we focus on in Netbox Labs? Putting in place guardrails that enable agents to do autonomous work when it makes sense, but require human-in-the-loop approval for most meaningful changes. And we have tools like Netbox Change Management that, you know, provide these guardrails. Agents today. So review and approval processes of changes that agents want to make to your infrastructure. Audit trails, super, super important. I promise you, you’re never going to be able to adopt AI agents in an enterprise without providing auditability, the ability to kind of look back and ask, like, what did, what did this agent do?

Kris Beevers • 16:36

Who approved it? Why? You know, what, what was that trail? And so branching and change management workflows, change logs, audit logs, these kinds of mechanisms are really going to be important. as you adopt agents day to day. And then lastly, validation tools. One of the ways we can adopt agents and be confident adopting agents is if we can prove to ourselves that the change being driven by agents in our infrastructure is in fact accurate.

Kris Beevers • 17:08

Like we’re not, you know, blowing up the network configuration, for example. And so for us, tooling like our discovery and assurance products, which is about finding drift in the environment versus intent, is really, really critical for unlocking confidence in driving a higher rate of change in the infrastructure, which is what’s gonna happen as we adopt agentic workflows. So, you know, it’s really important to think in a disciplined way about how do we create like safe frameworks within which agents can operate so that we can get the leverage from these kinds of tools. Now, I’ve talked quite a bit, but just wanted everyone to have a sense of how we think about all this stuff at Netbox Labs and some of what we’ve learned working, you know, across our whole ecosystem with teams that are adopting agentic operations workflows. And now I want to transition it over to John, who I think is gonna, you know, take us through some of the same grounding or some aligned grounding and then get us right to the fun stuff, which is, you know, real agentic workflows. So John, I’ll go here and hand it to you.

John Capobianco • 18:12

Well, thanks, Kris. And yes, there actually is going to be a lot of synergy and overlap, particularly around— you see governed and secure on my very first slide, deterministic execution, Ready by Design, a lot of the same themes that Kris has talked about. So in case you’re not familiar with iTential, we’ve been around for over 10 years doing— focused on orchestrating your network automation and being an orchestration platform. But now we’re evolving and adapting to become an agentic operations platform with the lessons learned from the years of doing network automation. Right, so we are the platform that will deliver that governance, the scale, and the trust required to orchestrate these agents safely. By deterministic execution, what I think is really clever that iTential has done is they’ve made all of that 10 years of history.

John Capobianco • 19:09

If you’re an existing customer, if you’re looking just to make an agent have access to your Ansible playbooks or your OpenTofu, your Terraform, or just Python scripts, we’re just adding the capability to the platform to make these tools accessible to AI agents. I just recently did an experiment around the SSO and RBAC where we provisioned a Flow AI agent with an API key with access to an OpenClaw system, and OpenClaw was not able to run certain workflows in iTential because of our RBAC system, right? So we were really— that was one of the very first things we tried was governance around how something like OpenClaw might interact with iTential. You can go ahead, Kris. So this is really where I believe there’s a rhyme with network automation and the journey people went through, where the first few steps are going to be humans in the loop, where the agent is asking permission of a human in the loop to say, yes, I approve that, it looks good, go ahead, or no, I need you to revise this setting or this change. And show me your new version of the change.

John Capobianco • 20:27

We’re going to be in the loop and there’s going to be experimentation. That’s where we’re all going to start with. And a lot of my work really does start with just experimentation because it is a brand new technology. It’s a brand new way of working. It’s a new way of interfacing with technology. But I think right as you— once you’re done experimenting or part of your experimentation, should be MCP servers and still having human in the loop, but now we’re going to connect That LLM, that model.

John Capobianco • 20:59

Now say it backwards with me. It’s a protocol. And when you think of protocol, think of email, think of websites, think of HTTP, think of BGP, think of OSPF, think protocols. But this protocol is responsible for the context and the model, right? And the model could be large language, small language, multimodal, right? There’s many different forms and shapes of models, but the key is that context in the middle.

John Capobianco • 21:29

So you want to provide Netbox, for example, the context from Netbox to your prompt. And it’s easy. Within a minute or two, you’re going to be able to snap in the Netbox MCP server It was my Hello World MCP server. One of the first ones I wrote was a NetBox MCP server to try MCP and to abstract the 130 APIs in NetBox. Not to abstract them, but to interface with NetBox in a new way through natural language. But this is going to move rapidly to humans on the loop where you’re maybe getting, you know, reports, or maybe you’re just designing and you’re staying on top of this thing and monitoring it and kind of treating it like a pet still where the humans are on the loop. But we’re going to be building and deploying agents moving beyond that human in the loop.

John Capobianco • 22:29

The problem with network automation, right, you have this awesome Python script or PyTS job or Ansible playbook or something, for a while you’re going to have human operators running those and monitoring those and managing them. But how do we move to full autonomy? That’s what the last thing in this chart shows you. Once we’ve orchestrated some of these agents with humans on the loop, can we get humans out of the loop or humans in the lead, let’s say, where the agents, and like I’m going to show you today, have full autonomy, and think of them as reasoning and acting agents, react agents. The reasoning capability came out with O1, the model from OpenAI, and the tool calling capability is the action. And those tools go all the way back to the MCP integration. All right, so we can move ahead, but that’s sort of the journey I see people doing.

John Capobianco • 23:25

And, um, so this is iTential’s operating model. We talked about operating models, right? In the center is the, the traditional iTential platform, and in the summer of this year we’re going to be releasing the, the Flow AI reasoning layer on top of that. And then we have an integration layer at the bottom, and as you can see, we’re going to have the Flow MCP gateway connecting via the NetBox MCP. So a full 3-layer stack where we’re gonna design automations in natural language, which are goal-oriented and persona-based agents, but they’re gonna leverage the same deterministic workflows from that platform, blending, you know, agency and static workflows to maintain governance and security and access control, further pushing down into a bring-your-own-MCP, a bring your own model, a bring your own agent approach because we’re agnostic. We want to connect and play with everything and everyone, and we believe that MCP is going to be that centralized protocol, like I mentioned, for us to interconnect these different tools together. And we’re going to be watching Netbox today.

John Capobianco • 24:40

Um, yeah, we can go ahead. So think of iTential sitting right in the middle of everything, right? That’s where we want to be. We want to be an agnostic layer, right? And if we look to the left, we see sources of truth— the Netbox Labs or Netbox. We can do Netbox Labs for discovery. On the right, we have the CI/CD pipelines that we can integrate with.

John Capobianco • 25:05

And the AIOps and telemetry with tools like Selector and others. Southbound is all of your infrastructure in whatever silo you have, your clouds, your data centers, your SD-WANs, your campus and security, and your classic network services. And then northbound are where we’re gonna integrate applications and agents, ordering systems, event and AIOps, and that SRE DevOps role at the top. So this is where iTential sits, is in the middle. And think of it— this is why I wanted to show this slide— is that an agent in the middle, if you think of that iTential logo as an AI agent, can call upon any of these boxes in any, any direction right, via MCP or REST API or even XML. We, we don’t care about the data format, right?

John Capobianco • 26:00

So we’re going to sit in that middle. But where Netbox really comes is that source of truth where we do want to orchestrate some of the southbound services. Well, we want to do that autonomously. We don’t want to have a human in the loop verifying the IP address or the site or the rack or the circuit, right? So that is where Netbox becomes a critical component of us enabling the agents to act autonomously. Yeah, go ahead. And so let’s switch over.

John Capobianco • 26:30

I’m going to share my screen right now. Share my entire screen. Oh, hang on. Do I have to restart? I should have tested this. I think I have to. I’ll be right back.

John Capobianco • 26:46

I’ve got to quit and reopen my Zoom.

Kris Beevers • 26:49

Always a Zoom thing. Always.

John Capobianco • 26:51

We’ll see you in a minute, John. Yeah. You’re going to let me share? Okay. No, there we go. All right. Okay.

John Capobianco • 26:57

So we’ve talked about the Netbox Labs MCP server. Here’s where to find it. Now, I want to say to the world, right? And I’m not just saying this because I’m on with Netbox and Kris is here. This is likely the best hello world introduction to MCPs that you’re going to find for a few different reasons. It’s 3 different tools which have read capability. And they’re very clear about this in the documentation that this is a simply read-only MCP, interact with your data via LLMs.

John Capobianco • 27:29

Now MCP, because it’s a protocol, is agnostic. So you can plug this into— and they give you an example of how to plug it into Claude, right? You can plug it into, um, Copilot in VS Code. You can plug it into Gemini CLI. You can plug it into anything. It’s, it’s universal, right?

John Capobianco • 27:48

So you need your URL, an API token, and away you go with how to run this in Cloud Desktop or even, you know, some of the other tools I’ve mentioned, right? So I’ve written my own version that has a few more tools. You know, and I did this because I wanted the CRUD activity. I wanted to be able to create and delete and read. But back to the official one, if you’re looking to work with your management or your leadership and not do shadow AI, Bring this forward and say, listen, I would like permission to use an approved LLM and connect my tools to my Netbox instances using natural language via the MCP. Now, the other MCP we’re going to see today is PyETS, which is my first MCP.

John Capobianco • 28:39

It’s been out for a while. It has a lot of stars, but I just wanted to show you this is how we’re going to drive and connect the network automation So in the Netbox community, and Kris, actually, I’m going to get you to jump in here. This is a bit of a legacy way, and there’s a little bit of risk using this site. You can actually provide a link to everyone on how to reserve an official free Netbox lab instance, right?

Kris Beevers • 29:05

I will do that. I’ll drop something in the chat for everybody. If you’re not familiar, Netbox Cloud does have a free plan. You can spin up your own. free Netbox cloud instance anytime you want. Works perfectly with MCP, so I’ll drop that in the chat. Thank you.

John Capobianco • 29:19

Okay, wonderful. So what you’re going to see is I have a device, and this device has 44 interfaces, and each of these interfaces are associated— most of them have IP addresses with them. Well, guess what? My agent created this. So we’re looking at the Flow AI dashboard, and what we’re going to do is we’re going to look We’ll work backwards. I’m gonna show you the mission log outcome, and then I’m gonna show you how the agent was built. So if we go into my mission logs, there’s one here, sync to Netbox.

John Capobianco • 29:53

Okay, and here’s the activity feed, which I’ll come back to. But the objective was quite simple, was to use the, the demo Netbox with this site ID and this device ID and gather the show IP interface and other information from the device and do a bulk create in Netbox and create a report with all links to the resources and a summary of what was found and what was created. So here’s the conclusion of what happened. It did find all of these devices and information, and it actually created all of the interfaces and subinterfaces and provided links to the resources in NetBox. all of the IP addresses that it created and the interfaces that they’re associated with, right? And this report could have been emailed to me or sent to me in Slack. It skipped a couple interfaces for various reasons.

John Capobianco • 30:57

Some interfaces were disabled. So here’s the mission logs. Again, the, the system input and the user input From that natural language, it went ahead and started to gather the show commands. Okay. And it did this all in parallel. These 3 commands ran in parallel. Right now, excellent.

John Capobianco • 31:22

I’ve gathered all my data sources simultaneously. Now let’s look up the Netbox ID. It used the Netbox MCP to look up the object. It used the bulk create to, um, Sorry, it’s a little bit of lag there. The bulk create to create the objects, and then that report that we looked at. Now all of this took 173 seconds, so, you know, 3-4 minutes.

John Capobianco • 31:56

And if we go to the manage agents, I’m going to show you how exactly this one was built. So if I go to my sync to Netbox, right, and I’m just going to hit the edit button so we can see how this was built. I have my system prompt, which is the system persona, and then I have my user prompt, which are more instructional-based, right? This was it. That’s all I did. There was no Python, no Ansible. Like, I didn’t write a lick of REST code or requests.anything, just these two prompts.

John Capobianco • 32:30

I picked my LLM provider, and then I bound it to the tools. So I bound it to the show run command, show running config command from PyETS, and the Netbox bulk create and the Netbox list objects tools. That was it. I hit run and it successfully ran the agent. Now this is syncing to Netbox, so I thought what we could do is do conf t here. int loopback, uh, let’s do 200, description of live demo, an IP address of just some random IP address, and we’ll no shut it. And now let’s just run the agent.

John Capobianco • 33:29

And I might just let this run in the background while we explore the other agent for the sake of time. But let’s just wait and see what it does. We’ll give it a head start. Again, it’s only 3 minutes, 3 or 4 minutes. It feels a lot longer waiting, watching it start, but it does only take a few minutes. And hopefully in the NetBox instance here, as you can see right now, there is no loopback 200. There’s loopback 100.

John Capobianco • 33:57

Right, so it’s going to call the 3 tools. And it’s running those PyETSTools serially or in parallel. And then it should move on to checking the netbox for the discrepancy. And ideally, hopefully adding that new loopback interface into NetBox. Let’s just let this finish up. I can’t see the chat while I’m sharing, so I’m not sure if anyone’s had any questions along the way or need me to clarify anything. I think we’re doing a pretty good job so far, John.

John Capobianco • 34:43

Great. It’s listing the objects. It’s probably doing the differential now, I hope. I just have to be patient. We’re at 90 seconds now. Okay, so there’s the device ID. I’ll create all of them in one shot, including non-deleted.

John Capobianco • 35:09

Uh, it’s trying to create all of them, and so maybe I should have adjusted my, um, my prompt actually. To say only add the, only add the, the deltas, right? Maybe I should have done that. I didn’t check my prompt. So let’s just cancel this and adjust the wording on that prompt. So this was sync to Netbox prompt. And we’ll just add it to be gather the current interfaces from Netbox first and only add missing interfaces.

John Capobianco • 36:02

So that’s how easy it is to adjust and to correct because the original purpose of this agent didn’t have any notion of deltas. The other agent, we’ll let that run in the background. I wanted to show you was I did the opposite direction. This loopback 100, I created it in the dashboard of Netbox manually after I synchronized. I made them out of sync. We have another agent in here. Yes, let’s just let it refresh.

John Capobianco • 36:43

Which synchronizes in the other direction. So when I ran the other agent, it has a slightly different toolkit. Let’s just get into the manage agents here just a second. It has a slightly different toolkit in that it has a PyTS configure tool. From the PyETSMCP. So if I go to sync from Netbox, the prompt is a little different. Collect the information from the device, collect the information from Netbox, right, and configure anything missing on the device that’s in Netbox.

John Capobianco • 37:26

So I’m still running the same Netbox tool, list objects. But now I’ve added the configured device. And if I go to the mission of that sync from and show you the mission logs, It’s quite interesting. So I, um, it’s right here, sync from Netbox. And here’s the conclusion: loopback 100 is up and up with the IP address specified in Netbox. So it went to the source of truth, it found this loopback in Netbox that wasn’t configured on the device.

John Capobianco • 38:06

And went ahead and used PI ETS to configure that device, configure that interface. So now let’s just check our current logs here. Our agent mission, we have one running in the background. Okay, it’s still running. Okay, check this out. Loopback 200 created. And now I’m checking the other IPs, and there’s that dummy IP that I had.

John Capobianco • 38:37

So if we go into NetBox here and refresh, there’s loopback 200, um, and there’s the IP address. Thank you to the demo gods all around the universe who got me through that live today. There’s a lot to talk about here. There’s a lot of implications here. I really wanted to show a practical usage of an AI agent vis-à-vis Netbox MCPs and the iTential platform. I think one of the biggest hurdles with Netbox is not using it, it’s populating it with your day zero information. How do I onboard Netbox?

John Capobianco • 39:22

How do I get all of those interfaces, all those IPs, all that information populated in a reasonable amount of time with a high degree of fidelity? We just did it today in less than 2 minutes. Now, how would that scale? Well, if we were using PyTS, it’s just scaling the devices in your testbed, or if you already have pre-populated all of the individual device names through a REST API called the NetBox. It can discover and match your hostname with the device name and NetBox in a site. There’s many ways for this to synchronize based on certain keys, and we did it in both directions. So that day-to-day maintenance of, you know, user Sally added a new loopback on the router, right?

John Capobianco • 40:08

The tedium of keeping these things synchronized is no longer a problem. We have an agent that autonomously— and I would build those two purpose-driven agents. And I would give them full autonomy to synchronize in both directions. Now, we would have to make the choice Is the device the source of truth or is the source of reality or is the Netbox? Because where do we want our users to be adding things? We want them to add them to Netbox.

John Capobianco • 40:32

Well, then that’s a pull and a push config and no one ever touches the network again, Kris. Imagine that. If all they’re doing is adding records to Netbox and the agents are synchronizing and pushing the IPs and setting up the interfaces and doing whatever, a true source of truth, a true intent-based setup. I see a couple of hands up. Did I miss any questions? Does anyone have anything? Shital, Dave?

Kris Beevers • 40:59

I’ll give just a bit of a voiceover on some of the topics that have flowed by, John, as you went through the demos. I think one of the themes that I heard, and maybe you can speak to this a little bit, is like, How do I get started tinkering with this stuff? So maybe you can talk to that a little. I have some thoughts too.

John Capobianco • 41:24

So we are in a vibe, you know, it’s funny, I don’t even know that people call it vibe coding anymore because that’s the way everyone codes. Everyone does it that way. So there’s no shame in starting a session in, in let’s say Cloud Code, Gemini CLI, Copilot in VS Code, whatever you want. Copy the link from, from Kris’s, uh, Netbox MCP and just ask it, I would like to connect this Netbox MCP to this ecosystem. Here’s the JSON payload, here’s my API key, help me do this. Use the AI to help you use AI, right? Like, it’s, it’s kind of, I know it’s cyclical, but if you use AI to help you get started as a beginner Right within a few minutes, you’ll be able to, to use Copilot to talk to Netbox, for example.

John Capobianco • 42:14

Right. And then go from there. Could you please make a table of all of my circuits in Atlanta? And if you’re in VS Code, it will make a CSV file for you.

Kris Beevers • 42:25

Yeah, I just, I just want to strongly agree, John, with what you said there. The best way to get started with this stuff is just an experiment, right? And like, my favorite example is what you just described. It’s, you know, download Cloud Code, get it running, ask it how to connect the Netbox MCP server. It will help you do that and then start to interact with your, your infrastructure data. It’s safe to do that. John, you pointed this out earlier.

Kris Beevers • 42:56

It’s safe to do that because the official Netbox MCP server is a read-only server. It’s not gonna, it’s not gonna break anything, right? And, you know, when you’re working with an agent, something like a Claude Code, you get a real feel for how these things can work iteratively in, you know, what you described, kind of the, the react loop using the tools to achieve an outcome. And then if you project that all the way forward to what you just showed us in your demo, you know, that sort of experimentation is going to give you a sense for how you might apply a platform like Flow AI or another agent platform in practice, right? You know, you need to get get a real feel for that yourself. So there’s some other questions flowing, John, that I wanna surface. I think there’s some folks in the audience who are pretty excited about NetClaw.

Kris Beevers • 43:45

So maybe you can speak to NetClaw for a few minutes. I mean, we all hear about these OpenClaw or like NVIDIA’s big announcement at GTC this week was NemoClaw. You’ve got NetClaw. What is that and how should people engage with it?

John Capobianco • 44:03

Oh, wow. Thank you. So I’ve done a lot of open source work in my life, and I’ve never seen such a reaction to a project of mine. It crossed 300 stars in just a few weeks. I see people using it in the wild. Someone just in the chat says it’s fantastic. So I took the core OpenClaw idea.

John Capobianco • 44:23

Now, OpenClaw, think of it as an assistant. And what makes it to me, what distinguishes it or makes it a differentiator, is that it comes preloaded like a battery’s included, almost like a Django approach. where all you do is fill out some simple settings and now it’s connected to Slack or it’s connected to Teams or it’s connected to WhatsApp or Telegraph or like it’s very, very simple. You put in your API keys and away you go. And now you have this personal assistant and not through the hype of things, but I just wanted to say, okay, let me try to build one of these, but with a network engineering persona. And add the skills. It really is skills-driven, which are Markdown files.

John Capobianco • 45:06

You yourself can build one of these claws quite easily, and it relies on the CLI and APIs and MCP. But Netbox is part of it. I added that Netbox MCP that I showed you as part of NetClaw. We added it to my VibeOps forum or our VibeOps forum. There’s actually a channel where this NetClaw is live and active. We’ve seen it do everything from configure very difficult network configs, make DrawIO images, NetBox, ServiceNow, you name it, it can probably do it. I’m really excited to see that people are talking about it in the community, and it does work with iTential.

John Capobianco • 45:50

I mentioned that we had the NetClaw agent in Slack interacting with the Flow AI agent in iTential. So we had sort of A-to-A communication working where the two agents were working together, and I thought that was quite remarkable and respecting the RBAC and respecting the governance. So yeah, it’s quite the— and the number of stars that OpenClaw— I mean, it has more than Linux. It has more than Linux on GitHub right now. So I would encourage people to, again, safely and prudently— don’t do this on your production networks. Make sure you’re working with the right management and security folks. I don’t want anyone to get into trouble.

John Capobianco • 46:31

But if you can find a sandbox area or a home network or maybe get approval to run this in a lab or something, I’m more than happy to help you. So reach out and let me know how I can help.

Kris Beevers • 46:42

Yeah, I really appreciate that, John. And just in case it helps anybody, you know, I think one of the best ways to get to know where the world is headed is to use these projects like OpenClaw or NetClaw or NemoClaw or whatever claw on your own experimentally. In a safe way. The future is going to look more like these tools. They’re not, you know, a few folks are asking about kind of enterprise environments and enterprise readiness. They’re not generally baked for that yet, but you can bet that that’s where a lot of investment is going. And even, you know, the Flow AI platform that John, you showed us today is really about that.

Kris Beevers • 47:26

Like how do we start to orchestrate agents in real production workflows safely. There’s a couple other questions bouncing around that kind of, touch on this theme. And so one tactical one that Tanner asked is about secrets and keys and credentials. And like, how do you think about, you know, enabling your agents to actually go connect to devices in the network safely without handing them the keys to the castle?

John Capobianco • 47:57

That’s a really good point. I think as long as you are still doing good key management and applying best practices, whether you’re using a secrets manager PyTS has something called secret strings that we use to encrypt the PyTS testbed. There’s a variety of ways, but yes, security needs to be top of mind, especially in this world of CLAWS and especially in this world of autonomous agents. What I found very interesting, NetClaw in the Slack channel came under attack by about 35 different social engineering attempts. People were trying to get it to do everything from divulge its secrets, divulge its API keys, change the boot var on the router so that it wouldn’t reboot. People were trying to do some pretty malicious things against it, but the guardrails baked into an OpenClaw or a NetClaw actually prevented it from doing any real damage. So I think that

John Capobianco • 48:55

It’s kind of like very similar. Again, it rhymes with network automation. Network automation forced a lot of us to learn about how to manage secrets, right? We’re network engineers and typically we have PuTTY or something like that. And now suddenly we’re doing CI/CD pipelines, we’re doing infrastructure as code, we’ve got Ansible inventory files, we’ve got PyTS testbeds. So it forced a lot of network engineers to pay attention to that. So repeat those best practices, right, in this world of AI.

Kris Beevers • 49:26

Yeah, I think, I think really one of the ways that I think about AI and agents for us as engineers is that it lowers the barrier to automation. But all those same principles, you know, really do map in an agentic world. Here we’ve got agents doing automation, but that’s not a lot different than having, you know, deterministic or orchestration code driving automation. We still need to care about secrets, we still need to care about guardrails, we still need review and approval processes for changes with a high blast radius, that kind of thing. And you know, we talked about this at the beginning of today’s session, guardrails, audit trails, you know, those sort of human-in-the-loop validation processes. That stuff is going to be really, really important as we ask these agents to take on more. Another theme that’s coming up that still is in this realm of like how best to get started, you know, I think there’s a lot of interest in

Kris Beevers • 50:23

local LLMs, which models are best to work with and build these agents. Maybe you can speak to that for just a few minutes, John.

John Capobianco • 50:32

Right. So I think that I love that there’s an open source element to the modern landscape of AI because it drives innovation, it keeps costs down. If there was no open source alternative, who knows what the monthly fee might be for something like ChatGPT. And for people that maybe are privacy-minded or security-minded or want to run it in an air-gapped environment, certain production systems, a lot of enterprises are reluctant to connect to some new cloud for an LLM. So there are 3 tool frameworks I suggest people look into, and they all have feel and look a little different. One is Ollama, and it’s sort of the de facto tool for this. Not Llama the model, but Ollama.

John Capobianco • 51:18

Before Kris and I are done this conversation, you could have downloaded Ollama locally and pulled a model and start asking it questions locally for free offline before the next 10 minutes is up in this conversation. Now, I’m not up on the latest models. However, there are big players in this space. Google has Gemma 3, their open source model you can run on Ollama. Microsoft has the Phi series, Phi 4 is their latest. There’s Mistral.

John Capobianco • 51:51

There’s NeMoTron from NVIDIA. I’m going to miss some and I feel terrible, but there is a wide burgeoning market for open source models that will run on a variety of hardware, Kris. You can get 7 billion parameter model, 8 billion parameter model, all the way up to 100 billion parameter model. Then the other two I would mention are LM Studio, which has more of a GUI look and feel and maybe is more beginner-friendly, I would say, than Ollama. Then also Microsoft has Foundry Local. Now the thing that for more advanced people listening to this, if you haven’t started doing it locally, all three of those platforms offer a REST API. Meaning all that code you see people pointing at ChatGPT, you can just swap out ChatGPT for Ollama in one line, right?

John Capobianco • 52:41

And all the code will still work. So you can do some really, really awesome stuff for free, privately, and locally with Ollama and local models.

Kris Beevers • 52:51

Yeah, I want to take a slightly different tack on this because, like, having now done a bunch of webinars on AI in infrastructure and AI in operations. This is always a topic and I understand why, but I actually want to challenge everyone who thinks that they cannot use the foundation models, the Claude’s or the OpenAI GPT models or whatever it is. It’s exciting to be able to experiment with these local models, but you will not really feel what the future is going to be unless you’re working with the bleeding-edge foundation models, which are mind-blowingly good at this point too. So, you know, find a way. It costs a little bit of dollars, right? But $20 a month or something like that, if you’re not spending it, like, that’s among the best $20 a month that you will, you’ll ever spend to, you know, I think experiment with the bleeding edge as well.

John Capobianco • 53:52

I totally agree. And don’t I wouldn’t judge the current, right? There’s a difference here. An open source model, as wonderful as they are, they are nowhere near as good as some of the latest bleeding edge commercial models from the hyperscalers. And one thing I wanna mention, Kris, before maybe we got about 5 minutes here. I would pay attention, everyone out there, to a new way of developing solutions with artificial intelligence beyond vibe coding, this spec-driven development.

John Capobianco • 54:24

Kris, have you been doing much spec-driven development? Of course.

Kris Beevers • 54:27

Yeah, absolutely.

John Capobianco • 54:29

Yeah. I would look into this, everyone. If you look into spec-driven development, SDD, if anyone’s ever done test-driven development, it will feel a little bit familiar to you. It rhymes with that. But there’s a SpecGPT or SpecKit from GitHub that you install into your Claude code and it offers slash commands. So /speckit constitution and it makes the constitution for you for your project. Anyway, take a look at this.

John Capobianco • 54:58

I’ll be having a video coming out soon about it. But I just, in terms of what’s on the horizon, Kris, I would say spec-driven development’s probably on the horizon for me.

Kris Beevers • 55:07

I just want to comment on that actually. You know, I think one of the things that, we can look to as a community, network engineering or infrastructure engineering community, is what is happening in software development. It’s a leading indicator of what’s going to happen for us. And in software development, we’re seeing models like spec-driven development or the importance of guardrails and evals and validation mechanisms. That’s what really unlocks trustworthy code generation by these agents. And I think we’re going to see the same thing in infrastructure management.

Kris Beevers • 55:42

We’re going to see Again, like safety mechanisms, guardrails, and validation tools give us the confidence to lean into and enable agents to operate against our infrastructure. And then how do we dictate what the agents should do? Intent and spec-driven development is not very different than clearly defining your intent. And so you can imagine that really resonates with me, John.

John Capobianco • 56:08

Yeah, I think so. And I honestly, Kris, I think that your platform in the very near future is going to be a platform where you’re going to have agents, the agent information, agent metadata, right? The Netbox of the future may also be a place to have a source of record and a source of truth about your actual agents that we build, right? So I think it’s a really exciting time. I want to thank everyone who joined us. I think it peaked at like 270 people today, so really incredible turnout.

John Capobianco • 56:38

Thank you so much for just for deciding to spend some time with us here today and ask questions. I hope we got to everything.

Kris Beevers • 56:46

We got to most of it. And yeah, I mean, first of all, thank, thank you all as well from me and from the Netbox Labs team. Just a reminder, we will send the recording, we’ll send the slides, and I mean, you can find us on Slacks all over the place, everywhere, on Discords, or, you know, all kinds of, all kinds of places. So I think both John and I were happy to engage elsewhere as well. And thank you, John, for taking the time with us and for some great demos today. Really appreciate it. Okay, and I think we’re going to wrap here, so we’ll close this down.

Kris Beevers • 57:21

Thank you all. Take care.