Enabling Agentic AI and Securing the Data Perimeter with Opengear SMF and UDOP

For those of you that don't know open gear. And as Tom mentioned, we've been, uh, doing this for, what, 13 years here. It's about a 20-year-old plus organization.

We focus on IT and network management solutions. And about 20 years ago, founded to give remote serial console access. And we've just grown beyond that point.

And we tend to use Tech Field Day as a way to come in, kind of give a peek behind the curtain as some things that we're thinking about, things that we're working on. It's all grounded on the open gear platform. We build on top of production.

What you're gonna see today is building about three different AI agents to do different things. That's all on top of existing running code. About nine-ish weeks ago, I changed my focus and if you guys remember anyone that was here last year that looked at the open gear presentation, we had the sess come in.

And that's 'cause myself and my team were really heads down doing some operational things and getting some things out the door. And as we've kind of transitioned beyond that, I've come back, kind of poked my head up, so to speak, and started looking around and the whole AI thing was out there, right? I mean, everybody hears the buzzword, et cetera, but I wasn't really paying that same level of attention to it, um, that I probably really should have been.

And when I took a look around, it was very eyeopening and frankly a little bit terrifying, I gotta say. And you know, he was working with Andy on who will be up here in a bit to, uh, come up and, uh, work the technical side of the demo and with one of the AI assisted IDs, right? Coding tools.

And one of them we tried, we weren't getting great results. He tries another one. I get a message from him on Sunday that goes, everyone should have this.

About 15 minutes later, I get another message that says we're all gonna be out of our jobs. Um, now what we've actually found out is someone with that level of experience who's, you know, been a chief architect, you know, layer two through what, seven aspirant to places like that, it just makes them way more efficient. Probably 10 x more productive in reality.

Um, which does open up a lot of questions about, you know, what happens with very junior people? How do you get up the career ladder? You see a lot of that, um, in the press today as well.

And so one of the things that always runs through my head with something like this is the old REM song. It's the end of the world as we know it, and I feel fine. And it's like, yeah, I don't know how fine I actually feel here, um, with some of, uh, what we're starting to see because I do think there's gonna be some big societal impacts, uh, as a result.

You look at the business process outsourcing firms in India, you know, for example, a lot of what I was just talking about can replace those, um, kind of lower level programmers that are just kind of writing code as well call center operations outta the Philippines. So there's definitely those things are all, all really happening. Um, but as we look at, you know, how do you make an agent or do you go with it, one of the things they need, of course is contextual data and how do you get contextual data?

And as we started playing around with these more in a network sense, um, looking at how we could actually do that, how we could facilitate that kind of building off of the open gear platform at the end of the day. And because we connect into, you know, virtually anything that could be sensors, that could be, you know, a serial port on a switch that could be, you know, a management port on a server, right? So all of those different things give us the ability to pull data in, um, very contextual data as well as push control back out in certain instances.

So this is my nod to Las Vegas, by the way, right? So, you know, what, uh, happens in Vegas stays in data, uh, in Vegas. What happens, you know, in a silo stays in a silo.

And so much of what we do is businesses is we tend to put, uh, you know, our data into different systems. And that system has a different login. It has different access controls.

You know, you've got a jungle of different ticketing systems, you know, logs go one place, something else goes another place. And all of that ends up, you know, in a different, uh, data silo. And at the end of the day, I think it was Nvidia that turns this an AI factory, right?

You gotta be able to take the data in, you basically process it, um, and you get intelligence out the other side. Well, if you're only feeding it something that's coming out of one of those silos, you know, call it security or maybe it's coming off the kind of OT side of the house, right? With some kind of an ops thing, then that's all you're going to get out.

You'll end up with a very, you know, siloed agent that can do a very specific thing. And I think, you know, from an agent perspective itself, we don't want 'em to be, you know, kind of the omnipresent thing that's going to be able to take over. Obviously, you know, the more narrow their capability set is, the less risk associated with them.

But you can't get to true intelligence until you really break those, uh, silos down. And it's one of the things that, you know, this kind of new platform starts to lead to. It was earlier this year, and I don't know how many of you have seen this, but it was in February, um, that Microsoft's CEO came out and said, SAS is dead.

All SaaS application is, it's a fancy UI sitting on top of some static business logic that interacts with your data, performs a CRUD operation. And he's like, it's dead. And about four-ish weeks ago in a podcast interview, uh, mark Benioff really kind of acknowledged the same thing.

Um, he was talking about their existing SaaS applications and then he got into agent force and how this is going to enable, you know, digital labor, which is gonna be this multi-trillion dollar market, right? So he is kind of painting that picture, but really saying the same thing. And I think if you look around, you know, in the networking arena, there's a ton of different SA applications that do different things, whether that's a D sim, you know, to different observ, uh, observation type, uh, tools.

Um, they're all really SaaS applications. At the end of the day, you've got a kind of static layer of business logic with some kind of fancy UI that sits up on top of it. Um, and I think, you know, this will break out different ways.

If you look at something like Salesforce, they got a ton of, uh, data regarding, you know, customers in some way that could be your CRM data from a sales perspective could be support Now. And they may end up kind of looking like a walled garden, almost like the apple of ag agentic ai. They take that data, they kind of consume that inside something that Cisco could do.

Take all your apps dynamic, your Splunk, all of that data, that's a great set take. Your CCIE program is a way to basically say, this is the knowledge this thing should have. And you could make a digital CCIE, you could rent that great model, but again, how useful is that without your contextual data of what that means coming off of your infrastructure, whether that's at the, you know, facilities layer or whether that's coming up the stack.

So looking at what happens as SaaS is dead. And on the left side of the slide is your SaaS application. And you know, every good SaaS product manager has a metric, it's called engagement.

It's to keep you engaged in that platform. They're measuring, you know, how much clicking are you basically doing? And so you can look at these things, it's almost a way to keep you engaged like that and you feel productive, um, as a result and sit on top away you go, well, if you replace that with an agent, what happens?

The agent's sit, sit on top. But now you've got this new data perimeter that's exposed, and that's really where all of the proprietary content of a business is. It's in your domain knowledge.

And if, how do you secure that? How do you enable that? And those two key things I think are super important to being able to really enable, you know, agentic or, you know, digital labor, uh, if you will.

And so how do you build that perimeter security? How do you give it the context that it needs without allowing leakage? Because if you allow that to leak, then your proprietary, you know, differentiating data just leaked as well.

So I'm gonna step back for just a second, and I, most of you, uh, in the room, I would assume know what an agent, uh, an agent is. There's really several generations of these things. You know, we were laughing earlier at, um, whatever the, the agent had responded as far as some tasting notes, uh, as well.

But you know, your general GTP type interface is an agent. It's just a Gen one agent. You go in there, you human, you put a query into it, it's a one and done.

You could go in and get, go, Hey, gimme a sample config for, you know, ISIS for this type of thing. It'll spit something out. Um, no real context beyond that.

You'll get into your gen two agents. This is maybe a depiction, um, a little bit more of what, what a gen two could look like because you're starting to bring in, um, other things, right? You're bringing in outside data retrieval, you're putting outside tools into play that human in the feedback loop that can help control, uh, decisions.

And with a gen two agent, they're a lot more kind of like a agentic assistance. And they're the kind of like I would see inside of a, the IDE type tools where you can put together a framework, you can have it do kind of pick and shovel work, if you will, to develop software code for you. And then you get to gen three.

And if you kinda read in the common industry, I think these will start to appear probably somewhere around 2027. And that's really where the agent, if you will, is working in that outer loop. You can give it a very generic description to go do something and then it can take action.

And that's really where you need to be able to bring all the different contextual data, uh, together as well. And you can make those workflows kind of as complex as, uh, you need to do 'em to hook different things together. Most of the agents we're showing in here today, um, are fairly simple, uh, where we're taking some amount of input, um, you know, vectorizing that putting that into a different large language model.

And one of the things we found, we were having a bit of a discussion before this started, uh, different large language models will give you different results. It's kind of like horses for courses. Um, we found certain ones, you know, will work better at analyzing log data.

Certain ones are better at, uh, doing other tasks. And so you can kind of manipulate that as needed from a, uh, uh, from an agent perspective. So looking back to, okay, what would it actually take to get one of those gen three agents to work?

The first thing we were talking about is being able to pull all of your data in, pull that into some kind of a store, get it in the right format, um, put access control onto it so the agent can get there. Um, basically that context firewall, if you wanna call it that the agent can now has the knowledge to do something, but then you need to access control. Again.

If you had built a, some kind of an agent, let's say, to do fp and a analysis for you, that thing should not have access to storage, shouldn't have access to network. And so, you know, just like different roles that you have, um, inside of an organization, these things will have to have the same thing. And then you need that control loop or that feedback loop.

And I think this is where, you know, the world of, uh, simulation, digital twins, et cetera, can sit in there to help really train what's going to happen on the other side. Now, on the left side of this equation, in November of 24, philanthropic released, uh, MCP or the model context protocol. And what it is built to do is to try to take disparate data sources and normalize those so that a agent can ingest them.

So back to my finance, uh, example, if I wanted to pull something, you know, from public financial data, I could make an MCP server, put that in front of Yahoo, uh, finance, pull that now in, I could get some kind of report out the other side. So any place that you would want to, uh, pull data from you would put an in MCP server. And it's really just a, you know, client's, um, server type protocol.

They didn't take security into consideration when they first developed this. Now they're working on it, they're working on it quickly and hard, but there are some difficult problems that still need to be addressed. They're not simple ones, which is how do you determine the identity in a distributed system of each one of those MCP servers so that you can't put a rogue server in there and inject, uh, you know, fake context or whatever it is to corrupt the whole systems.

That's one of the keys right now. The protocol itself is, you know, not encrypted in flight either, so you've gotta, um, deal with, uh, getting that encrypted. So there's a number of things like that, uh, that they'll be working through and they are working on, uh, you know, kind of as a, as a larger community.

About the same time Google put out their agent to agent, uh, protocol to be able to get an ecosystem of agents work together. There's some super interesting concepts inside of it. Um, which is your really your agent capability card.

It's almost like a resume, if you wanted to use that anana analogy. Here's the skill sets that I have as an agent. Of course that allows, um, you to be able to use those to be able to put other controls on.

Um, so there's a lot that's going on from a protocol perspective, but to really build what we need, I do think we need really a new platform to be able to pull this off. And this is a little bit more, um, you know, kind of IT and OT specific, but I think it could make it more generic for, uh, business in general. You know, in this particular case that lower level down there is a place where, you know, we as open gear have historically played, whether that be sensors, whether that be telemetry or whether that be the management side of this.

Now Prototypically, you know, coming in from serial console, it's on the far right of that, where the console server would go in, that's great connected. We can pull data out, we can push control back down. Um, but likewise, we do pull telemetry up from our connected devices and oftentimes sensors are, uh, in there as well.

And that could be power, that could be environmental, um, you know, dry sensor door, open door close, those kinds of things. And that next layer up, um, or that independent data management path is what we call the smart management fabric. It open gear.

And so it's a way to connect all of that. It's kind of a, an overlay of fabric, if you will, from a management perspective. And then being able to put controls on top of that to say that this particular, um, entity can only do these things down through, uh, the management fabric database on top.

What you'll see today is what we call the connected resource catalog, um, to be able to, to, uh, get all of your inventory, data, et cetera. And then sitting on top of that is really your identity management layer, uh, that's coming in and then plugging agents on top of it. And you know what it, what it does is it allows you to secure that data perimeter, that contextual perimeter.

It, um, gives you the ability to put some governance around it, on what agent can do, what, what human can do, what, what level of access can you get, uh, where, and you know, those are the things that we need to do, I think to really enable AI operations, uh, at the end of the day. Um, with that, what I'm gonna do is we're gonna turn this over to the, uh, to the technical side of the, the demo with that, that intro and, uh, come up. So, uh, Andrew Pierce or Andy Pierce is our chief architect, and Matt Whitmer is, uh, a solution engineering for us.

And like I said, uh, earlier, we tend to use these as a way to get some feedback on things that we're looking at going forward, exact, you know, dates, et cetera. So it's a bit of a view into Our roadmap. The product management team could get into, um, you know, offline if they want to as far as, you know, dates or when something like that, uh, could come into, come into play.

But it is built, you know, a hundred percent on existing, uh, open gear platform running, you know, production code. We're just really using sometimes what I call our a agentic UX to be able to, uh, do this stuff on top of, and it really comes back to, I think, core open gear. It's in the name of the organization that openness, that ability to get into, uh, the solution to extract things or to push things into it so that you're not dealing with a closed system in some way.

So.