72. Networks Need Agentic AI with HPE Juniper Networking – Tech Field Day Podcast Spotlight Series

AI is taking the world by storm, but are we ready to turn our networks over to a software program to tell us the best way that we should be doing things? Is there a better solution? Or has complexity finally conquered our ability to model and manage networking?

In this episode of the Tech Field, a podcast networks need AgTech ai. Welcome to the Tech Field Day podcast, where we bring together a group of influential IT technical experts to discuss single idea or premise about topics in enterprise it. This podcast features a variety of perspectives from members of the Tech Field Day delegate community, and is often associated with one of our events.

Tech Field Day is a part of the Future Room Group, and this podcast is also published on our sister company's website at Techstrong tv. In this episode, brought to you by HPE Juniper Networks, we're gonna be talking about networking and agentic ai. But before we get to that, I'd like to take a moment for our guest to introduce themselves so you know who we're talking to, starting with Keith.

Hi, my name's Keith Parsons. I run a company called Wireless Line Professionals, and we do wifi. I've been doing it for o over two decades now, and I'm glad to be here as part of MFD.

Alright, and Sunol. Hi everybody. Super excited to be here.

This is Sun VP Products, uh, as part of the now new combined HPE Juniper HP networking business unit. I'm here to talk to Tom and Keith about, uh, simplifying network operations with some really cool new tech. And I'm Tom Hollingsworth event lead here at Tech Field Day.

Let's jump into the premise for this episode. No doubt you have heard all about how AI is going to change the world, and there's some big picture ideas out there about what we're using it for, but we still have to operate the things that AI runs on top of. And that means networking.

What if we had some kind of thing that could help us make our networks better? What if there was something out there that could give us all the information we needed and reduce the amount of time it took to diagnose problems? Well, that's not a what if anymore because a agentic AI is here.

And in this episode, we are going to debate the premise that networks need a agentic ai. Now, I've been doing Tech Field Day for a very long time and I've seen a lot of things come and go, but I can honestly say that I've never seen more discussion around AI than I have well ever really. But AI seems to be a hot topic that is really capturing people's attention because it is a very visible representation of what we want a computer to do, right?

We, we want the Star Trek model, we want the computer to answer a question and give us information and do something. And historically we haven't been able to do that. There's a lot of typing involved, a lot of head scratching, a lot of, well, I didn't think it was gonna do that.

So I, I kind of want to turn this over to, to Keith and Cini. What is it about AI that makes people feel so excited about where we are going in the future? We've been on the AIOps journey for simplifying network operations for the last 10 years, and we have seen it give visible benefits to network operators.

And I think the premise of does AI as a tool help simplify network operations has been proven in the sense of when issues happen and we know issues always happen. Like Keith, you mentioned, right? You've been in the wifi industry since forever, the problems haven't changed.

People still say wifi sucks, but the question still remains, how do I solve that problem? Right? How do I know why wifi sucks?

And it could be a wifi problem, could be a wired problem, could be a van problem, a user device problem, or an application problem. But how do you know? There are various ways that we have tried to demystify why wifi sucks.

It could be anything in the network stack. AI seems to be helping us accelerate that journey of how can we understand what is causing this bad experience and agent take AI seems to be the new, uh, kid on the block or the new shiny toy that can really help us accelerate that journey to truly identifying user experience problems and maybe even one day automatically fixing, fixing them with self-driving. Well, I, I have a comment and I'd like Sun Lady to come and answer it.

I've been attending Juniper missed presentations for a long time and, and, and, and love the AI journey. And just lately in the last round of presentations you've been using this term ag agentic ai, and I, and, and the first time I heard it, I had to go look it up in the middle of the presentation just to even know what you're talking about. So maybe our audience doesn't really know what ag agentic AI means.

Could you tell us the difference between that and what we've had before? Oh, absolutely. I think that's, that's a great topic to dive deeper into.

So agentic AI essentially is a framework where you're using artificial intelligence for reasoning, reflection, thinking, and recommendations, right? And that is exactly today what a human does, whether it's a personal situation or a big problem in networking, right? We are looking at all the information as we know, we talked about our AI journey, the insights, the data, and then we reflect on the data, we analyze the data, then we reason the data.

Why is the data saying what it's saying? What is the conclusions I can draw? And then we come up with a conclusion that says, oh, here's the root cause of why there was a problem in the network impacting users.

That's exactly what the agentic AI framework, when we put it in a networking paradigm comes to the fore to communicate, right? Because now the thought process is what we were doing with standard A IML, what we were doing with reinforcement learning, with unsupervised machine learning, with supervised machine learning, with deep learning with various models. Can I now leverage LLMs?

Can I now leverage agents interacting with LLMs in this sort of non-linear programming language to automatically get the data, analyze the data, reason, reflect, think, and come back with a recommendation as to what is causing a bad your experience, right? So agent take care essentially is nothing but a conglomeration of agents, each agent doing a certain aspect of the problem solving statement, how we then take it in and apply it to networking because again, the problems haven't changed, just the ways to solve it are changing every month, every quarter, every year. That is where agent take AI comes in to help us solve problems much faster and potentially even resolve them, uh, automatically.

So you bring up a good point and it's some, I wanna go back to something that you had said when you discussed this originally that I I latched onto. And it's this idea that we hear that the problems are all the same and, and you used a very good general one. The wifi sucks.

That's what we always hear, right? I I would posit that that's not a problem, that's a symptom. The problems change behind the scenes all the time to influence what the symptom is.

And I've been doing this job not as long as Keith, but I've been doing it long enough to remember that in the old days it was bad frame relay configurations and rip routing poisoning issues. And now we're doing completely different networking stuff on that side, and you have multiple different spectrums that you're trying to analyze and, and things like that. And the easiest analog that I can think of for this is if you've ever had a problem with your car and you've gone to Google and you tried to Google what that problem is, you know, very quickly you have to be very specific about it.

You have to put the year and the model of your car in, because sometimes you'll get really weird results like, oh, well your carburetor jets are out of alignment, but you have to know that your car doesn't have a carburetor because it's not that old. What, I guess my question is, does egen AI have enough intelligence to kind of take a look at that knowledge base and say, okay, I can immediately eliminate these 18 problems as being the root cause because these aren't configured or we don't do these things anymore. And it it, it prevents people from getting locked into a, a troubleshooting, um, set that will not produce an outcome because there's no way you will ever be able to configure these things on a version of software that's like 19 years old at this point.

Yeah, I think that's a, that's a great point, Tom, because this concern that just brought up of vi take me in the right direction or the wrong direction, actually also came up about two years ago when JI first came to the fore, right? When GPT sort of changed the world December, 2023, everybody was going to chat GPT to ask the, the most silly question to the most serious question. And sometimes you got the right answer, sometimes you got the wrong answer.

Similarly, when you take that, you know, two years, fast forward the agent a KI there is always, uh, a risk of hallucination, right? Because the LLMs know what they know, but they don't know what they don't know, right? And therefore, it's not a simple lift and shift of, hey, if I create a bunch of agents and expose them to an LLM, life is good and all my problems are solved.

They're not, as you rightly said, networking. The symptoms are the same. It all comes down to wifi sucks or my apps not working.

But behind the scenes, there's a lot more complexity in what that network entails. 'cause apps are now sitting in the data center, apps are sitting, uh, SaaS applications. I have multiple van circuits with the whole SD van play of people moving from MPLS to, uh, to multiple, you know, circuits.

So there is a lot more data that needs to be analyzed to be able to figure out what is causing that symptom of wifi sucks. And that is where just creating agents is not the be all end all end of the day. You have to make sure that the efficacy's high.

So that always some guardrails around the agent KI framework. So the levers you have are the right prompt, like you said, right? When you're looking about solving an issue with a car, you have to give it the right model, the right year to get a more specific answer that's more relevant to you versus general car problems.

Exactly the same way from a networking domain perspective, there has to be some domain expertise and some feedback built in to make sure that when the agent is reasoning, reflecting and thinking it is giving you the more relevant specific conclusions versus here is my API, here is my MCP server, and I'm gonna use a new term here. The the MCP concept that's also taking the world by storm recently, it's part of the new automation, but just having a model context protocol where users can now interact with all my APIs via agents, is not going to really help move the needle with high efficacy. For high efficacy.

You need a human in the loop, you need some guardrail, you need to train the agents to look for exactly what type of data based on the prompt. And that is what leads to the right answer. I I have a question on agentic ai from a, uh, a personal point of view.

Many of us have been doing networking for a long time and how is that gonna affect us in our jobs and our careers? Is it gonna take over? I mean, if we go back years, Cisco used to sell something and say it's a network engineer in a box and it it'll do it all for you.

And, and we learned that that didn't happen, is is it about time that this is actually going to be a network engineer in a box? That's a very interesting debate, Keith, right? There is, uh, one point of view that says that, um, support is no longer a necessary function because AI will do that support for you, right?

That being said, it really cannot solve everything, right? Taking it back to the concept that Tom mentioned earlier about is ai, the be all end all, you know, do we, is it gonna give me the right answers all the time? It, it won't, it needs guard rail, it needs training, it needs human feedback to make sure it's always being course corrected, right?

So will it relegate all of us and take away our jobs? My viewpoint? I don't think so.

What will make us do though is become more productive and more efficient to get to that root cause much faster. So what took us hours and hours and sometimes days and days and weeks and weeks will now take us maybe minutes to get to the right answer, and therefore we can get to operating more networks, operating more networks more efficiently, and creating more applications on that network for users to use. So it's not so much a, a bane of our existence, it becomes a boon because we can now do more right?

And more of the right things of innovating and creating more services versus doing the sustaining aspect of networks, which is where most of the time goes today when customers running their networks. What are your thoughts on that? Oh, no, I I, I, I agree.

I, I see it as a, as a benefit. Um, but I, but I wonder if it has any of the downsides that humans bring to networking and that, uh, we had this experience years ago and that we remember it, and that little piece of knowledge can be incredibly useful in solving this problem or could cause us to go down the wrong path. Because last time I did this, and, and that's not the problem today, as humans, we, we have these memories of when, uh, Tom used to work at Gateway and he had this one thing and it comes back and you had to set our IRQ.

Well, like you said, we don't have set IQs anymore, but does that as a human, sometimes we go down the wrong path. Well, having AI do this, make it so we go down less wrong paths, is, is there less of a chance for that personality to, to bring that kind of crazy memory in? Uh, great point.

So let's take a a look at how AI operates, right? At another very high level framework. End of the day, the AI is as good as the data it's fed right now, if you look at what data it's fed, it's fed data in terms of the actual network operations.

So what is the telemetry coming for? Every user, every minute across wallet, wireless, across wired, across wan, it's also looking at all the data that's publicly available on all the issues found in the past, right? Specific to, let's say a Juniper network or a Cisco network or, or whatever the case might be.

So if the data is there and there's enough weightage to the right data points, the right elements vis-a-vis those, those, uh, offshoots of memory like you just mentioned, you know, in the human case there is a law of large numbers efficacy that will come into play that will always correct it in the right direction versus taking us in a hallucinatory path, right? So it all comes down to the right data. For example, one of the key things you've heard us talk about here at hp, Juniper is all about AI driven support, right?

So in addition to the network operations, the network configuration, the telemetry coming for every user every minute. So one of the key things we have to look at is the ability for us to ingest that support ticket data, right? The whole concept of Marvis efficacy that you've heard about before from Mist Juniper has been about feeding it the information that comes from our customer tickets and seeing if Marvis can answer the question and the Marvis cannot answer the question we go into saying, Hey, did Marvis have the data but did not have the signature to identify that same problem?

Or did Marvis not have the data at all? Therefore, I need to get right the right data into cloud. But NetNet, by looking at the customer network itself, the telemetry coming from the customer net network and looking at the corpus of support tickets you've already had, the whole goal here is that AI has substantive information to make it go in the right direction versus feeding off wanting of thought, which, you know, our US humans sometimes suffer from and go down troubleshooting that path.

It has a much larger data corpus to look at to give us the right direction of how to solve a problem. So it all comes down to the right data. So, So I wanna jump in here and ask the question.

Based on what we've heard so far, using eent AI to solve existing problems seems like a no-brainer, right? We, it will speed time to resolution, it will ensure that we're, we're doing the right thinking around things. But what happens when age agentic AI runs into a problem that it doesn't know anything about it, it this is a new issue or it's so cutting edge that, that we've never seen this before.

How can age agentic AI help speed time for a human to reach a resolution, not knowing anything about what's going on? And that's a great point, and again, I'll go back to the framework for agent a KI because the, the symptoms have not changed. Like you said, the root cause may keep on changing, but oftentimes, and not the root cause also seems to be in the same way, right?

There is a loop in the network, there's an empty U mismatch negotiation failed bad cable, you know, authentication issues, things like that abound across, uh, network operations. But if there is a use case that KI has not come across before example, there's no support ticket for it and therefore there's no data for it to react to. That is where what it'll still do is it'll, with the various agents that you develop as part of this framework, one agent called another in a non-linear fashion and tries to at least gather all the data.

So its reasoning and conclusion may be not in the not as high efficacy as we want, but it'll still be able to present the user, Hey, here's everything I know about this problem based on the information I have access to. But it, it may still require a human, the loop to go say, okay, this is what the actual problem was based on this data, and that is what the agent I will learn from. Because now when it knows this new problem came in, I could get the data for it, but I couldn't really get to the right root cause because the customer said, eh, you did not gimme the right answer.

That's fine. But that's where when you ask the agent to now relearn, uh, and retry that question or that query, it'll give you the right answer. And that's again, another benefit of ATech.

KI, whether it's a reasoning and reflection. If it gives, if it reasons the wrong answer and you ask it to reflect on it, it'll then come back with a better answer. 'cause it knows what it did in the past was the wrong thing.

So from the human's point of view, when that failed, what did the human receive to give it help? Other than that's just, does human just say that's wrong, try again? Or does it give the information, uh, about that problem?

E Exactly. So what a human would do in that case is the levers we have, again to get the right answer is the prompt, right? You engineer the prompt in such a manner.

So when you ask it the first time, you give it a prompt, it gives you an answer, but the answer's not quite up to par, the answer could be just flat out wrong or the answer could be incomplete or in the answer could be great. In the case of the answer being incomplete, you essentially say you give it a thumbs down and say, Hey, go reflect again because this was not fully answered. And it's gonna go back into all the data, look at other areas of data which were a match to the query, and try and come back with more responses.

In the case of when it's absolutely wrong, you essentially tell it, I'm giving you a thumbs down, this was not the right answer. Go redo the whole analysis all over again. And that at that point in time is gonna take a complete different approach.

So the agents that we're interacting with the reflection agent, the reasoning agent, the analysis agent will go retrigger that whole cycle again to retwe itself to say, okay, what I did in the past does not quite work. Here is my new answer based on the whole new medication, that cycle with a different prompt, right? In the day, it's the data and it's the prompt engineering that we do that helps the NLM engine and the agent take KI agents, so to say, reason in the right way.

So, but the one key part that, that I also wanna touch upon is part of the agent take KI, right? So far we've been talking about how can agent take AI frameworks and you know, this plethora of agents that are interacting with each other and reasoning by themselves and reflecting, analyzing and coming up with recommendations. How can I use them to actually sell, drive a network, right?

Like you, Tom, you talked about your car example again, you know, self-driving cars are reality now, right? With Waymo and, and Teslas and everything. Several networks are now a reality too because what we are talking about is not just being able to find the issue based on a prompt.

Imagine a world where based on all the signals coming into the network, that the agents can automatically autonomously find an issue and automatically autonomously if you've given it permission, fix the issue for you, right? We've been doing this already, as you know, in the juniper missed framework with Marvis actions where, you know, with the port stuck, we'll bounce support whether when it's a a firmware noncompliant issue, we will operate the firmware for you if you give us permission. So there is a trust aspect as well because obviously as much as we all love self driving networks, we also want to make sure there's a human the loop to give permission to trust the maus action to take the full self-driving loop.

But the key part is as we add more and more such self-driving actions today, we are looking at the self-driving being all about the network itself. For example, if I have a missing vlan, if I have ant U mismatch, if I have a lie detected issue, can I go shut off the port? Can I go change the configuration and add the VLAN automatically?

Because I've already found the issue That is, again, where agents could come into play and make the, make us enable this whole self-driving framework much faster. But think about in the world of agents, even in across the entire networking stack, right? You have, uh, DHCP, you have authorization, you have application servers.

If the, the Marvis agent finds an issue that is leading to a conclusion of saying it's a, it's a problem on the DH DHCP server side, wouldn't it be nice if it, if the pharmacy agent could interact with the d HCP server agent and say, Hey, go fix this issue or increase the number of leases that you have because your DCP pool is already, uh, you know, uh, consumed and therefore user are not able to get an IP address, and that could be truly self-driving end-to-end across the network, right? So that is again, where we are seeing a paradigm shift that this technology could enable where not just within the, the net the core networking stack itself, but even beyond the network when adjacencies cause issues, can agents interact with third party agents and sell, drive the whole network stack end to end, right? That's the really exciting part.

So You just extended that level of trust outside of the network and say, oh, not only do I trust my agentic agents to do their job within the network, I also trusted to go talk to the SaaS server or to the outside servers. Um, I, I like the idea, but that that's, that's a another trust level that you have to be really con confident before you turn that part on. Absolutely.

And, and that's what I think, again, going back to your point of will as humans still have a job, I think we still will because we still have to give it that permission to say, yes, I trust you to be able to go do this, right? I trust you within my network domain. Do I trust you beyond it?

Do I trust you to interact with a third party system to enable this? So, and is this the right conclusion? So that aspect I think will always be there of human and AI always interacting together.

I I do worry that this process, it is fixing the, the sup bottom layer support gets fixed very easily, but where the humans that started in support, many of the great engineers I know today started in support and that job is basically gonna be gone. So how will we build the next generation of network engineers when they don't have that, that baseline to start from has nothing to do with this topic. It just came out.

So Tom, what do you think? Well, My curious teeth brings up a really good point here that you, there's a level of trust that a lot of people have with the way that things are done. And we've already seen that network operations and engineering teams are cautiously optimistic about trusting AI to automatically do things.

And, and, and there's also, there's levels in there of like, here's the suggested fix. Do you wanna do it? Versus I did the fix.

Do you wanna look at what I did and, and back it off if necessary, but then it's the idea that these systems are talking to each other because one of the things that we've seen over the years is this territorial IT problem, right? Where well, DHCP is technically a server function, so it belongs to the server team and you guys can't touch the servers without permission and your boss has to talk to my boss and, and you're effectively saying, we're gonna eliminate that permission chain, if you will, by having the systems talk to each other, which in theory fixes a lot of problems, right? In practice though, your technological advances bump into organizational issues, if you will.

And I can remember when we first started talking about automation, this was the whole thing of, oh, you're gonna try to produce, deploy those, uh, changes in live without a maintenance window, without people sitting over your shoulder making sure everything's gonna work. Oh, that'll never fly. And we've slowly gotten to the point where we, we've trusted the systems enough to do that.

Do you feel like there's gonna need to be a policy discussion at a higher level to help IT teams understand that agents are doing what they think is best and that by creating layers of policy approval above that you're potentially impacting reliability and, and those kinds of things. I think those concerns are definitely valid because you are right. I mean the organizations when there are different domains of control, it always comes down to you're not touching my system.

You know, even though you are saying my system is a problem, you always have to prove why that's a problem. So processes and policies will change where they will allow for more and more automation to self-drive the entire end-to-end spectrum. However, it has to gain that level of trust.

And again, it all comes back down to the efficacy and the right data and the right model to give the right answer. And uh, you know, the feedback we've been getting from customers is when you know that an end user is in distress when their traffic is getting black hole, don't wait for me and my teams to come together and do a change control window because the customers or the end user already in distress, just fix it, right? If you know a camera feed is not going through and you know it's it's mission, you know, it's, it's very mission critical for me.

Um, bounce support. If you know that you're having issues on an end user where we know that it's caused because of a missing VLAN and the voice call is dropping and it's going into la la land and the voice call is not going through, go fix the missing vlan, right? These were not trust aspects that we got right on day one, right?

It took us three to four years of proving the efficacy and showing our customers, Hey, this is the issue that you, that you talked about. This is what Marus already found for you. And right now we are, you know, in the driver assist mode where we are notifying you, but more and more they're saying, if you know it's a problem, just go fix it.

Now that has definitely changed in the last two to four years in the networking domain itself for just campus and branch. But slowly but surely, Tom, I feel that this will happen across system that are even not controlled by the same organization, right? Like we talk about end-to-end assurance, we talk about from an end-to-end assurance perspective for end users when they are using application, be it a DCP service or some other that's sitting in the data center or a SaaS application and we know that's causing issues to the end user, how can I fix it end to end, right?

So right now where we are in campus and branch we self driving where when it's beyond campus and branch these adjacent services, we just going to drive our assist mode and say, Mr. Network operator or Mr. Network operator, here's the problem.

And please then escalate to the respective teams to go solve the problem with all the evidence that that problem can be solved. But I think surely as agent a KI evolved, and as more and more systems start supporting agents, you will see a more agent to agent interaction, right? Like what you and I were talking earlier when the session start started is agent TKI the new form of automation, it potentially is and it has a lot of scope given the reasoning and reflection beyond the manual workflows we would automate earlier.

Right now we're talking about automated workflows happening with agents and configurations getting fixed automatically. But yes, it'll take time, it'll take time to earn that trust, but we'll get there. As you can see, it's very exciting in the world of networking because we are making technological advances above and beyond anything we thought could possibly be a thing we needed to worry about.

We have software programs that are doing diagnosis for us and we are reducing the complexity on the people who are learning about how networks should operate. That to me is the definition of a paradigm shift. We're not doing the same things the same way repetitively over and over again.

We're doing things at a much different, faster, more exciting level. And I think honestly, a lot of the solutions that HP Juniper networking has been developing for years are the way that we're gonna make that happen. Sini, I know that people are probably very curious to hear about what you're doing as well as our good friend Marvis.

If people wanna learn a little bit more about that, where can they go? I'm very excited to talk about this virtual event we are doing in September. Please do come and visit us at AI native now the new era, um, on September 16th and 17th.

We will send out the details of registering for the event, but that's why you'll come and learn more about how we are leveraging Agent a KI to deliver self-driving networks and making sure the user experience is the best possible. And we'll make sure to include that link in the show notes as well as links to the recent presentations from HPE Juniper Networking from Mobility Field Day, and from other events. We want to thank you very much for listening to this episode of the Tech Field Day podcast.

If you enjoyed this discussion, please make sure you subscribe to our YouTube channel or in your favorite podcast application so you don't miss an episode and consider giving us a rating, a review, and even a comment about the topic that we discussed today. This podcast is brought to you by HP Juniper Networking and Tech Field, a, the Home for IT experts from across the enterprise, which is a part of the Futureum Group. com/podcast or find us on Techstrong tv.

Thanks for listening in and we will see you next week.