Open-Source Internet of Agents and a Quantum Network | Cisco’s Vijoy Pandey
Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, explains AGNTCY, the open-source Internet of Agents that Cisco built with Google, Dell, Oracle and Red Hat and has donated to the Linux Foundation. He describes how agents from different companies can discover, trust and work with each other, the secure messaging layer SLIM, and early real-world deployments. He also sets out Cisco's work on quantum networking.
Transcript
In the race to create the most accurate and trustworthy AI platform, it's not all cutthroat competition. Cisco has brought together some of the biggest names in the AI space, Dell, Google, Oracle, and Red Hat to create the Internet of Agents. AGNTCY has now graduated from concept to reality and is revolutionary for the future of AI interoperability.
But wait until you hear what Cisco is doing with this technology. Vijoy Pandey from Outshift by Cisco is talking with Agents of Tech today about this exciting new venture.
Welcome to Agents of Tech coming to you from our Washington DC studio. I'm Autria Godfrey joined by my co-host Stephen Horn and Laila Rizvi. And straight ahead today, we are unpacking how all these different AI agents now in existence might discover, trust, and work with each other across different companies and platforms. This is really exciting. AGNTCY now in existence and they're donating it. Can you believe it?
I think it's incredible. I mean, 2025 is the year of the agents. So for for a company like Cisco to be donating a network effectively of agents to something like the Linux Foundation is incredible, especially when things are just so cutthroat, so secretive in the field of AI that everyone is, you know, poaching talent from each other, spending billions of dollars to make even more billions of dollars. So, it's a completely different approach in that it's not a single model. I mean, it's a network and it's incredibly generous as well.
Very. You know, Stephen, it's interesting to think that, you know, all of these different companies, Dell, Oracle, uh, Google, they all have their own AI platform, but unless you're operating just solely on that platform, you're kind of dead in the water.
Right? It's really interesting, isn't it? Because two things that we don't often talk about, we don't talk about interoperability very often. It's not very sexy subject, right? But it makes sense, doesn't it? We have interoperability in other parts of life and these companies work together all the time. So why shouldn't they in the AI space? And we talk quite a lot about agents, but I don't think a lot of other people talk a lot about agents, and and that really is moving the needle, isn't it?
Yeah. It's nice to see these companies can all play nicely too.
Yeah. And I also think...
For the betterment of humanity.
Yeah. And I also think the way that the internet started, and Cisco was obviously heavily involved in that, was it was quite a nice thing which everyone contributed to, that you had things like MySpace, Bebo, as in it was a way to communicate with people. And so, but that snowballed and gave us all the data we needed to be able to get to where we are with AI. So I think it's looking forward and understanding, okay, what are the foundations, what are the, what's the infrastructure, the networks that we need to build to take us to the next stage? Because actually no one really foresaw what happened with the internet and where that would lead us technologically. So I think Cisco is...
Absolutely doing that in a way that maybe others aren't.
Right, they were kind of ahead of the game there. Well, you talked about communication. We of course want to keep in constant communication with you. Have you had any experiences with setbacks with interoperability issues when utilizing different agents? Be sure to weigh in in our comments, because the more feedback we get, the more we can tailor our conversation specifically to you. So now, let's get to it. Here now is Vijoy Pandey from Outshift by Cisco.
It is our pleasure to have Vijoy Pandey with us here today. Thanks so much for your time.
Thank you for having me.
Well, you serve as senior vice president and general manager of Outshift, what you have described as Cisco's incubation engine or hub, where you move with startup speed but enterprise resources. First off, can you kind of explain for us what that means with regard to AI?
Yeah, so if you think about Cisco, you think about a networking company first and foremost, and we were the, the one of the people who helped launch the original internet. So if you think about the internet, you think about Cisco, you think about the networking and routing equipment that we launched to enable that internet to happen. And if you think about the original internet, the whole goal of that internet was to accelerate global businesses. And if you think about what happened after that, the cloud internet happened, where we had mobile first applications and a social first economy happen at global scale, and Cisco participated in accelerating that movement as well.
If you think about agents and where AI agents are going, AI agents are here to solve for and accelerate all of human work. Whether it's in B2B, B2C, whether it's scientific or mathematical discovery, whether it's physical work in the form of robotics and embodied AI, or the way we interact socially in all of these social platforms that exist, agentic AI is going to interact and impact all of that. So if you think about that, we need a platform and an open interoperable internet to bring all of these facets together. And that's where we come in as a company.
And when we think about problems in the Outshift space, we think about three metrics for success. One is product technology fit. Is the technology mature enough to solve for these needs? The second one is product market fit. The technology is mature, but we want to bring a product to the market, and is it solving for business needs for our customers? But finally there's product company fit, which is whatever we are building, is it aligned to the go to market and the scale that a company like Cisco has? And when you think about the Internet of Agents, that collaboration platform is actually pretty aligned to where Cisco has been and where Cisco wants to go.
So the goal here is to let AI agents from different companies and platforms discover, trust, work with one another, just like you mentioned. Okay. So this sounds great, but it also sounds like that's a lot of agreement to collaborate among participants that typically compete. Is, is that correct? And so then if that is correct, then what does this actually look like in practice?
First of all, let's take a step back and figure out how this is all going to pan out and how this is going to work. And if you think about agents, they're getting very similar to what humans look like. So the way I describe agents is that they have humanlike attributes and characteristics, but they operate at machine speed and scale. And just like humans, we all come together in teams. We're all subject matter experts. We all come, come together, collaborate in teams to solve for bigger goals and bigger problems. Or some people might want to have agents behind walled gardens as monoliths in their own wells, so to speak. That's not how agents work. That's not how humans work. And that's not what the future looks like.
So what we believe, and the reason we picked the name Internet of Agents, is like, just like the original internet, it needs to be open and interoperable, because an open interoperable system drives maximal value for every entity in the ecosystem, whether it's vendors like us, whether it's operators, whether it's customers or developers. Everybody benefits through an open interoperable system. And if you really draw those parallels behind how a team comes together to collaborate in the human world and how agents come together to collaborate, there are lots of parallels here that we can decentralized and distribute.
Let's take an example. Let's take a movie that we're trying to create. And let's say you want to first discover talent. So you want to discover a director and a bunch of screenwriters and a bunch of actors. All of these people need to come together. They're all subject, subject matter experts. They need to come together to build a movie. First you need to discover them. Then you need to figure out whether they are authorized to work on a movie, whether they are who they claim to be. So there's an identity problem. So discovery, then identity.
Then you need to put them together in a team. So this is the movie, this is the goal, folks, we're solving for this goal. So that's the composability problem. How do you bring them together? And composability could be fixed like a workflow or it could be self-forming. And so that's composability. And then finally, you need to run with that movie and produce that movie. And finally, you need to look at whether they're producing what they're supposed to produce. So the evaluation piece comes in. Did you really do what you were supposed to do? Not just as an individual, but as part of this movie making team. And so if you think about this entire process and the movie making process, for example, it's a highly distributed process, and we draw a bunch of parallel from there to the Internet of Agents, and we've broken it down into those phases, and we're tackling each one of those phases in a highly decentralized and distributed manner.
But so let's zoom in on the evaluation and determinism cycle, because with deterministic software you can fix a bug, you know that it's solved. But with probabilistic agent-driven systems, it doesn't sound like it's that simple. So, is there a way for us to evaluate and know whether they are actually working as intended?
There are two parts of that problem, actually. So, there's a part that says, can you evaluate just a singular agent? Because an agent itself needs to do what they're supposed to do and it should not hallucinate. It should be on topic. And even beyond hallucinations, let's say we built an agent that solves for bank transactions, it should only solve for bank transaction. It should not go ahead and plan for your vacation, because agents are expensive and LLMs are expensive. So how do you keep things on topic? And so while you build the agent, can you evaluate the outcomes through the usage of things like existing workflows, existing outcomes, existing data that you feed from your environment, but also from generation of synthetic data? So based on the environment that exists around you, you can generate synthetic data and evaluate agents on that synthetic data.
Guardrails is the other big thing, where you want to make sure that it's not biased and it's not, uh, outputting things that you shouldn't be outputting, and it's not, u, it's secure. It's, it's, it's not, it's, it's immune to things like prompt injection attacks and it's immune to... I mean, agents are almost like new interns, and they are trying to learn things in your environment, and you know what can happen with a new intern. So how do you make sure that the intern has enough guardrails around them that they don't do something stupid? And so that's all of these things you need to do in the evaluation phase.
The one point I would make here, because we brought in determinism. What's happening today is we were always working with deterministic software, and with AI and generative AI specifically, but also with agentic software and auton, autonomous agentic software, we have probabilistic software in our tool chain and in a toolbox. So now we have deterministic software as well as probabilistic software in our toolbox. It does not mean that you take that hammer, which is the probabilistic software, and you go and search for a nail everywhere. We need to be really careful in how we construct these software systems moving forward, where, which have a combination of probabilistic software as well as deterministic software, and let's not try and force fit everything into the probabilistic domain. And that is really important, because they both have strengths and weaknesses, and we need to make sure that we are leveraging the strengths and weaknesses properly.
That's interesting. You keep talking, you know, about trust, and, you know, of course trust is paramount when it comes to the success of AI, because if we as users don't have complete confidence that the agents are providing accurate output, then where are we? So to that point, when agents from these different vendors all try to collaborate, there is often a kind of cognitive dissonance. So if they're not speaking the same language, how do you overcome that to create that trust, so that this collaboration is secure and reliable?
And so that, that is a big problem that the industry is trying to solve. And even though you'll see a little bit of noise saying, yes, we've solved this in a multi-provider, multi-framework way, what we've really solved for is the syntactic nature of that communication. So let's say you are speaking English and I'm speaking German. The way you construct your statements in English, noun before verb, might be different from the way I construct statements. And so that grammar manifests itself in agents based on where the agent has been created and where it sits. So whether it's on an Amazon Bedrock, whether it's on Google Cloud, whether it's a framework like LangChain or LlamaIndex that we're using on prem or on a cloud. So that's the grammar.
What the industry has solved for so far is making sure that everybody understands the grammar. So now everybody's speaking English, but we know that that's not the end, because when I say something, the intent could be very different based on my intonation, based on what, what I have in my head and how I phrase it. And the same words in English can be, can mean different things to different people. So the next step is to take that intent communication and figure out how to pass that and make it more deterministic.
So it's, it's kind of getting everyone to have the same definition of best. For example...
When a human says, "Agent, find me the best price for this flight." What does best mean? It's not deterministic. It's very fluid. Then the agent comes back and says, "Here you go. This is the best price, and I'm 90% confidence that this is the best price." What does that mean? Like, and what does 90% mean, and is the 90% same across a Salesforce or a Cisco or an Apple? So I think that's the starting point. But from there you get into reputation, because if somebody means 90% consistently, their reputation goes higher. If somebody says this is a task based on the communication and I authorization for this task, then their trust goes higher. And so from there, trust and identity and evaluation, all of that flows from that semantic layer. And that's what we all need to work towards. But it's an unsolved problem. And we are all working on that problem.
Okay. So you've created AGNTCY, this Internet of Agents, bringing together several of the biggest heavy hitters in the AI realm to do so. You've put this together. It's revolutionary. It's incredible. And then you just give it away. You've donated it to the Linux Foundation. Some might say, why, why would you take the open source route?
The reason why Cisco is in this space is because we were one of the big players in creating the original internet, and we feel that things have not really worked out that way as we've migrated towards the different other parts of the infrastructure stack that have taken shape. And so for the Internet of Agents, we really want that internet to be open and interoperable.
But we also want to drive an opinion as to how this should take shape. We want other players, again, we want players from the operator community, from the vendor community, and what good is a platform or an infrastructure if developers and customers are not happy about it? So we want developers and customers, all of these four personas, to come in and help us shape where the Internet of Agents goes. And even though we open sourced it back in March as a Cisco forward open-source project with LangChain and with Galileo, we wanted to bring it to the Linux Foundation, which is a neutral foundation for neutral governance. It's really credible. We've worked with them in the past. They have the wherewithal to grow the community and maintain the community. So which is something that we would want the Linux Foundation to do. It's not Cisco's job to grow that community and maintain the community.
And finally, we want some big names to be part of the journey, and it's always better for a project to be at a neutral place for a big name to be part of the journey. And so we formed it with Google, Red Hat, Oracle, Dell and us as formative members. And we actually are seeing a lot of interest from other big, uh, vendors and operators and customers as well who want to join. And the whole goal here is to build a common steering committee that is not Cisco-led, that is led by all these players, because we are, we are moving the ball forward as a, as a whole industry.
Let's talk real world deployment. Uh, do you have any concrete examples of how enterprises are already deploying this at scale?
Within the past six months, I mean, it's only been six to seven months, we've seen so many deployments, not just from Cisco working with, uh, those customers and partners, but also from other players where we seeing some real world deployments. And so the first one I want to, uh, talk about is this, uh, company called SoftServe that actually is in the video, that built something in the video surveillance space. So here you have a multi-agent system that is talking to each other in a video audio environment, but is also accessing old school, old world tools and databases and software to drive awareness in that organization if something goes wrong.
For example, these guys took the directory to build all of their agents and to make sure that they are discoverable in terms of capabilities, because they're also working with a whole host of partners that are contributing capabilities into the space. They then leverage this messaging layer that we built in AGNTCY called SLIM, stands for Secure Low Latency Interactive Messaging. But this messaging layer allows many to many communication between agents that is real time in nature, interactive in nature, which is critical for audio video agents. And you get a bonus if you use SLIM, you get postquantum cryptography. So really secure communication by default. So right from the get-go you're secure for time immemorial. So that's something that you get as a bonus. But these guys are using that to build their use case and they've already deployed it.
So we've actually built a network configuration and validation test pipeline working with some telecom providers and service providers that is already seeing reduction in human involvement by 10x across thousands and thousands of network devices. And at the same time, it's not just efficiency, at the same time is making every network configuration change in their environment error-free. So it's, is really accurate and it produces efficiency.
The other one I'll briefly talk about is we've open sourced this cloud AI practitioner or cloud AI engineer, is called CAIPE. Is Community AI Platform Engineer, but is a cloud platform engineer, and it's 20 plus agents working together, again bringing more than 40% of the use cases within Cisco are being, are being sent through this multi-agent application, and it's reducing time from hours to seconds on a whole bunch of tasks that we've, we've done in the past. So these are all real world examples that we've deployed and measured, and they're performant, they're accurate, and they're saving us time and helping us do things that we should be doing and not tackling these mundane tasks.
You know, you've mentioned the speed at which this all came to fruition, um, a couple of times. And I would just personally be interested to know kind of what happens when you go from the pilot projects to running production systems. Is there an eyes closed, fingers crossed moment where you just kind of hope it's all systems go, or is this set up so that you can kind of work out the kinks as you go, but go ahead and get these production systems into existence?
So that, that's the great part about actually building it in an open inter, interoperable system like an Internet of Agents, an AGNTCY, because what you enable is a platform layer which is distributed, not owned by a single entity, is not a closed environment. So you can come in and experiment. You can come in and see which things are working, which things are not working. You can steer the conversation, steer the technology to where it makes sense, and you can pick piece parts based on where you are in that maturity journey. Because that is the biggest thing that we're seeing in this journey, to your, to your point, is like people, some people are starting out, some people are way ahead, and so you need to meet folks where they are in their journey, and an open source-based environment platform allows you to do that. So you can start with something simple, like just a directory, like can I just discover agents, bring them together, and then you can layer on identity and trust and performance through SLIM and so on. So it allows you to do that, but it's also, it's giving us feedback, like, so if this thing is not working, help us fix it, like help us grow it, file a bug, file a documentation bug, come and collaborate with us.
Right, we're all working together here. Okay, I want to talk about something else that's very big that you've also got in the works, your next-gen quantum networking infrastructure to try to connect processors across great distances. Where does this technology stand right now, and what are the potential real world applications of getting this up and running?
We all saw that ChatGPT moment in the world of generative AI, where November 2022 ChatGPT comes around, and there's some folks who are ready and some folks who are not ready, but then everybody's embracing it, and it's a step function change for human productivity and all of human work. There is going to be a ChatGPT moment for quantum, and it's happening. Is going to happen, and it's all a matter of acceleration. And you would see in the AI space, you would see announcements every month, every year, every 6 months, and now you see announcements every week, literally every day in fact. And you're seeing something similar happen on the quantum side. We used to see step function changes every few years, every decade. Now you're seeing some of these step function changes happen in every month, every few weeks, and we're seeing that acceleration take place.
And for us in the quantum computing space, again, going back to Cisco and the product company fit, we are looking at quantum networking and quantum security as the two places where we want to innovate, and we want to enable the quantum internet. And so to us, what I would like to say is our goal is to enable the quantum internet to connect multiple quantum data centers together, and quantum sensors, which already are here and now.
But as we walk towards that goal, there are real world use cases of quantum networking in the classical world of today. Things like decision coordination, things like figuring out if there is some eavesdropper in your environment, things like making sure that clocks in these large cloud data centers are synchronized, or clocks for these large astronomical use cases are synchronized. I mean, there are whole bunch of use cases where you can leverage quantum technologies, quantum networking technologies in real world use cases today. And that's what excites us, because that is here and now. That's something that we can monetize, and that's something that we can move the ball forward even while we wait for the broader quantum space to evolve.
Right. Well, speaking of moving the ball forward, I want to have a crystal ball moment here and ask you, you know, five years out, when the Internet of Agents and the quantum safe networks are mature, how do you see this changing the way industries solve global challenges? And if I could put you on the spot, are you able to pinpoint perhaps the first global challenge that you would like to see tackled by this new era of artificial intelligence?
So I, I would say the whole goal here, through the Internet of Agents and through agentic AI, is to democratize intelligence, to make sure that intelligence is in everybody's hands, no matter where they are, no matter their social economic status. You have intelligence that anybody can gather today. So an expert in Johns Hopkins and the kind of healthcare that they provide should be available to anybody in Asia or Africa or anywhere else, even within the US, regardless of your social economic status or location. So that is the eventual goal for agentic AI in my mind. And what that allows us to do is be a little bit more proactive and be a little bit more higher order problems, solve a few more higher order problems and not get stuck at the mundane layer.
So we want to understand both ourselves better and the universe better. Those are the two goals that humanity is after. And can we do it in a very democratized, distributed way, with an abstraction layer that allows everybody to participate in the same way? And to me, that is the end goal. The specific things would be, first, start with the self.
So make us healthier, make us happier.
Yeah.
So get the drugs out, uh, cure the common cold, uh, get, get cancer out of the way, make us live longer, but then help us understand the universe, because that is the ultimate goal for all humans.
Right? Well, Vijoy Pandey, it is exquisite and, and incredible what you and your team at Cisco is doing, and we certainly thank you for joining us today to tell us all about it. Appreciate your time.
Thank you, and this was a great conversation.
All right. So really fascinating, uh, new technology that obviously they've unveiled, but also a really interesting conversation. I, as an entrepreneur, as a business owner, I would love your, your thoughts on, um, their philosophy of startup speed but enterprise resources and the way that they're thinking things through.
That's the nirvana, isn't it? That's what everybody wants. They want to sort of act like a startup, but they want the resources and the money to scale and build the systems, and usually that falls flat in its face, right? Because, because, you know, one thinks one way and the other thinks the other way. But no, in this case I think that's, you know, that's really true. And I think the point Laila was making at the beginning about, about infrastructure, about systems, about building that in this space, which is again is what they're doing, is, is, is kind of like really interesting. Again, not the sexiest of subjects, you don't read a lot about it, but I think is, is, is really important.
And the whole concept of AGNTCY. I mean, when I was talking to Vijoy, my thought was, I, I can't believe it's taken somebody this long to, to really come up with this. You would think that having all of these platforms that could work, you know, collaboratively with each other.
Yeah. And have agents that you can discover, that they can team up with each other, they can work together and get the job done. Whereas, I mean, you've got lots of people saying, "Use my agent." But I, I don't think you can actually get that much utility from a single agent. It's like saying a single employee. Yep. I'm going to run my entire company with one employee that's really good and should hopefully get... Well, it's not, is it? It's not.
But you're already starting to get this problem where every single software, you know, piece of software you use offers you agents. How many agents can you have? I mean, you can't have a cabal of hundreds of agents, can you? These people have got to get these agents to work together, haven't they?
Yeah. Exactly. And you can't have, oh, actually my agent can't see that because of these guidelines and these terms and conditions. So like, that completely causes friction in the entire system, and so you don't get anywhere and that technology is stunted. So I think it's absolutely critical.
And something that, you know, that they're working on with regard to quantum. I mean, that's, I think, going to be their next big thing.
I feel we need to do another episode on quantum. Don't you think?
Think Vijoy could come back...
Just to talk specifically about quantum.
I think so. Yeah, they're really, they're really investing, and it seems like really shifting their focus to...
Yeah. And laying the pipes for what will be a quantum future, but not necessarily trying to, you know, build the next quantum computer which is going to change everything. But they're anticipating that technological change, and they're, they're again doing what they do best, which is kind of building these networks, this infrastructure, and laying the groundwork for this future environment, which is going to be inevitable. It's not a matter of if, it's when.
I don't think the when is that far away now before, you know, a quantum computer. I mean, look, we've had the, we've had the theory since Einstein, right? And now I think we are getting to the point where we will see a quantum computer in the next two or three years. But more to the point, there'll be a lot of applications that are already running that will run on those computers. And so, you know, interesting time ahead.
Our thanks to Vijoy and to Cisco for that discussion. And remember to weigh in on the comments. Do you see AGNTCY helping solve your interoperability problems? Your thoughts on this new technology being donated to Linux? We want to hear all about it. And while you're there, make sure that you like and subscribe. Thanks so much for joining us today. We'll see you next time.