How the Internet of Cognition Could Transform AI | Vijoy Pandey
Vijoy Pandey, senior vice president and general manager of Outshift by Cisco, returns a year after introducing the Internet of Agents to talk to Autria Godfrey about what comes next. He says AI agents can now connect and communicate but still cannot think together, and sets out the Internet of Cognition, built on three layers of connectivity, coordination and expertise. He also gives an update on AGNTCY, Cisco's open-source infrastructure for agents.
Transcript
We sat down last year with Vijoy Pandey, Cisco's Senior Vice President and General Manager of Outshift by Cisco. He laid out Cisco's revolutionary Internet of Agents, a future in which AI agents could discover one another, connect, and communicate. Well, now, one year in, that technology is being used in the real world, but Vijoy believes that connection is only the beginning. He is back with us again today with an exciting new update to the Internet of Agents, the Internet of Cognition. What is it? How does it work? And will it transform how AI agents and humans interact and work together?
When we last spoke, the Internet of Agents was still a new and emerging idea. What has changed over the past year and where has it now been deployed?
Well, a lot has changed in the last year, and here's what a year buys you. It buys you core infrastructure. An idea has now become core infrastructure for a whole bunch of enterprises out there. So we started by donating AGNTCY to the Linux Foundation. That was when we were talking to you last year, and that was a promise. It was a promise to do something even bigger. Today, we find that we have more than 1.7 million downloads of AGNTCY code. This is our open-source ecosystem for the Internet of Agents, and we have more than 150 organizations that are contributing and building on this code.
And it's not just these code adoption numbers, but this infrastructure stack which is open and interoperable is powering many critical use cases in production in enterprises. We've published a few of these, we've talked about ServiceNow. In IT service management, we've worked with SoftServe. In video surveillance, we've worked with Swisscom and their phone networks. Digital Realty, healthcare, a whole bunch of these use cases that are in production today. So yes, a year ago, this was an idea. It was a happy thought. Today, it's core infrastructure that companies are building their businesses on.
And it has to be really, really rewarding on your end to see that thought come to fruition. But now that these AI agents can connect and communicate, they still can't think together. So what is the problem with that? In my mind I'm picturing two agents that are able to communicate with each other but perhaps they're speaking different languages? Is that an accurate example?
Yeah, I mean, that's just the beginning, Autria. So you can see, like you said, we've now made agents be able to talk to each other through the work that we've done with AGNTCY. There's some other open-source projects like MCP and Google's A2A. All of these open-source projects allow agents to then talk to each other, discover each other, and be observable. But here's a surprise. I mean, talking to each other is not the same as understanding each other. And you mentioned not being able to speak the same language. German and English or Japanese, I mean, that's the easy part. Translation is actually the easy part.
Even if both parties speak the same language, let's say English, think about a problem statement like this where you have two agents, there's a flight agent and a hotel agent. And I know it's a cliche example. But when you go to these two agents and say, "Book me the best flight to Paris and find me the best hotel nearby, and you have $1,000." This is something that we deal with every day as humans. Guess what happens? The flight agent is sitting there thinking, "Best price. Is that the best outcome? Shortest flight? Best time of day? Best rated airline? What is the best flight?" Once that's done, the second agent, which is the hotel agent, picks it up, same questions, like, "What is best? Is it the best branded hotel? Best amenities? Best location?"
But then you look at nearby. What does nearby mean? Is it near the airport? Is it the center of town? In fact, is it CDG or is it Orly? I mean, we don't even know that. And then, by the way, the budget, you have $1,000. Who decides how much to spend on hotel versus flight? What you'll end up with is business class tickets and rooms at The Ritz. And you don't want that, you'd want the total budget to be $1,000. So yes, we've made agents connect together, but they cannot think together.
So you've realized this problem of not being able to think together and you've coined this concept of the Internet of Cognition. What exactly is that? How does it make these agents think together and work together? And for people like me who are really, you know, the general population outside of the industry of technology, how is that beneficial to me?
Yeah, so the best way to think about this is think about human intelligence, and what happened with human intelligence is actually happening with agents right now. So for the first 300,000 years, humans became individually smart. So they became... Their brains became bigger and bigger. I mean, our brains became bigger and bigger. We learned how to work with fire, we learned how to work with tools, we scratched on walls and we hunted and we sowed and we used all these tools in a very intelligent way. But none of that intelligence actually spread. It actually remained confined within the inventor or within very close groups.
And about 70,000 years ago, language showed up, and language became our secret weapon. Language allowed us to share goals, share knowledge, and language allowed us to compound intelligence. And now it is the agent stock. So language was our secret weapon. We want language to be the agentic secret weapon. And that is exactly where the Internet of Cognition is going, giving agents that same weapon so that intelligence can compound, not just across multiple agents, but across multiple agent and human teams.
So if you go back to the travel example that we talked about just in the last question, we want them to agree on the shared goal that we need the best flight, this is the definition of best, we want the best hotel rooms, this is the definition of best, and this is a common goal of $1,000. It's not 1,000 to each, it's 1,000 to the both of you. And we want them to negotiate and coordinate on budgets and locations and flight types to then give me and you the best hotel and flights within the common goal of $1,000. Now that is scaling out intelligence. We continue to build bigger models and we will continue to build smarter agents, but we also want them to reason together and therefore be able to compound intelligence. And that's what the Internet of Cognition is doing.
So is this just an issue of retraining agents on every piece of data possible so that they're able to think on that data?
No, it actually goes beyond that. So we want both axes of intelligence to take shape. So right now we've been building bigger agents, bigger models, smarter agents. This is the equivalent of the human brain becoming bigger and smarter and learning how to work with tools. But then there's the horizontal axis, we want to scale out intelligence. And that is the equivalent of human teams, civilizations coming together, working together on a common problem. And that's what we're trying to build. So think of it as three layers. There's connectivity, there's coordination, and then there's expertise. So connectivity is what we are trying to do through AGNTCY, where agents can connect together. But coordination and expertise, those are the problems that have been missing so far, and that's what we are bringing to the table.
So let's talk today. Where does this technology stand today? How much of the technology already exists? What's still in the works? And can you give us any kind of a timeline?
Think of those three layers that I just talked about, connectivity, coordination, expertise. They're in various stages of maturity right now. Connectivity is real, it's here right now, it's actually been deployed in production. We've talked about AGNTCY, the 1.7 million downloads, the real use cases in healthcare and in service management and in networking and so on and so forth. Coordination is taking shape right now. So we have published the protocols. This is the equivalent of language that allows agents to coordinate and negotiate towards common goals, come to common understanding around things like terms like budgets and so on. I mean, that was an example.
And what we've seen, we've published these numbers, what we've seen is that without these protocols, agents, even two agents, are less than 20% aligned when you give them a problem statement. With those protocols, they go to be more than 90% aligned. Now, 100% is a pipe dream because even humans, we don't align 100% of the time.
Right.
I will take 90% any given day. So we've gone from 20% to 90% being aligned and we are moving into real world use cases and applications with this coordination layer right now as we speak. So you'll see a lot more around this in the next few months, but expertise is next.
And so when you think about expertise, that comes from two things. One is accumulative knowledge. So this is context, and everybody's talking about context, but this is context across multiple agents and humans. But also expertise includes guardrails and shortcuts. If I am a human expert in something, I know what's off limits and what can be done fast and what can be done, what is the most expeditious way of achieving that goal. So that expertise is not just shared knowledge, but it's guardrailed shared knowledge, what's off limits, and what can be done in an expeditious manner. And that is a fabric layer. That's something that we're building next once the coordination layer comes up.
So is there a timeline in place for when the Internet of Cognition could be operating at scale?
I would say this is a pretty fast-evolving field. This is something that we don't think of, evolution in the intelligence space, in quarters or years. This is months.
Okay.
So think of the coordination layer is happening right now and you'll see real progress in the next month, let's say.
Wow.
And the expertise layer in the next three to four months. That's where the world is headed. It's not even dog years, it's fly years, because these things are changing by the day, changing by the hour.
And Vijoy, I know in addition to, you know, what you're working on right now, what we've been discussing, that Outshift has so much going on. What can you tell us? Any updates on AGNTCY or more technology that you've got cooking?
Right, so we are Outshift, I run Outshift. This is the incubation engine for Cisco. We are looking at emerging technologies and how emerging tech can shape the businesses that Cisco wants to enter. And we try and reduce both technology risk and market risk when Cisco tries to enter that area. So as part of our charter, we are looking at super intelligence and how to scale out super intelligence. So a lot of the work that we are doing in the Internet of Agents, which is AGNTCY, that's A-G-N-T-C-Y. So you can go to AGNTCY.org to learn a lot more about this, and there are, like I said, more than 150 participating organizations there.
But there's also a lot of work that's happening on the Internet of Cognition that you can find on our website, as well as other frontier tech that is going to shape our future, things like quantum computing and what Cisco is trying to do in that space by connecting quantum computers together. So there's a lot that you can find at outshift.com and it's a space that you can go and learn a lot more. We are not pitching products, we're not selling products. We are an incubation engine, so it's a great learning resource to just go there and ramp up on everything.
That's really exciting. So I mean, you know, we had you here a year ago talking about the Internet of Agents, so hopefully we can have you back a year from now talking about the success of the Internet of Cognition.
Thank you so much, Autria. And yes, it's been an exciting time for us, and we strongly believe that connectivity, even though it's been built and scaled, connectivity is not cognition, and cognition is the challenge to solve to enable a future where agents and humans can coordinate, can build knowledge, can solve for something net new that nobody else singularly can solve for, but do it in a guardrailed, safe, and accelerated fashion.