チームとエージェントをまたぐ知能の協調法｜Cisco Outshift SVP Vijoy Pandey

Generative AI is transforming industries at an unprecedented pace. But as AI changes how you or your team work, one thing is clear. Your skills also need to evolve. At DataCamp, we offer everything you or your team need to adapt and thrive with AI. Whether it's business users looking to get the most out of ChatGPT and Copilot, or developers and data scientists looking to fine-tune models, you can learn the entire AI skills spectrum on DataCamp. Power your AI transformation today. Start learning at datacamp.com. The future is going to be multi-agent human teams that are going to solve really hard problems for humanity. Agents today can connect together. They cannot think together because they cannot do simple tasks as grounding, resolution, coordination, negotiation, and so on and so forth. That is the most important tech story of 2026 has been that AI agents finally started working consistently. However, when you try to get lots of agents to work together on a task,

it's still rarely successful. Today, we're going to discuss how you can coordinate intelligence between agents, more generally between teams of humans and agents. And even at three agents, what we found was three agents were not aligning on a simple goal more than 16% of the time. That's just terrible. And sometimes there is this bifurcation in the path where you say, let me go with agent capability sacrificing trust versus let me go with trusted agents even though I might have to sacrifice some capability. Our guest is Vijoy Pandey. He's Senior VP and GM at Outshift by Cisco. Vijoy leads Cisco's Department for Emerging Technologies spanning agendic AI, quantum networking and next generation infrastructure. He also oversees Cisco Research and co-chairs Hisco's Responsible AI Committee. So, let's find out how we coordinate intelligence. Hi, Pete Joy. Welcome to the show. Thank you, Rishi. Pleasure to be here. Yeah, great to have you here. So, I want to start off talking about the cognitive revolution.

So, you've likened the age of agents to the cognitive revolution, I guess, just like 70,000 years ago in humans. What's the peril here? If you think about human intelligence and the human advantage that we all have, Richie, it's not that we can just think. It's the fact that we can actually think together. And if you look at what's happening with agents right now, it's following a similar arc. About 300,000 years ago, we humans got smarter. Our brains became bigger. We learned how to work with tools. We learned how to work with fire. We carved, we dug, we sewed, and we created all these tools and we got really exports of them. but all of the innovation that humans did disappeared with the inventor they either stayed with the human or they stayed within pretty tight siloed groups and then about 70 000 years ago what is now known as the cognitive revolution took place and the reason it took place is because of a singular thing and that is the advent of language

so language came about and instead of intelligence being a solo sport it became a collective sport so now humans could go and say let's build that village together let's capture that hymn together so instead of doing things individually we moved towards collective intelligence and language allowed us to actually do three things first and foremost it allowed us to to align on intent, again go after the same goals the second thing was to share knowledge and this phrase that we hear all the time standing on the shoulders of giants is actually because of sharing knowledge is because of language because you could record things for posterity and you can compound knowledge over time and that's how civilizations got built and so these two things then allowed humans to reason collectively instead of reasoning individually and this entire arc is exactly what's playing out

in agents and this is what the internet of cognition is all about that's a very cool thing and yeah obviously like the last 70 000 years have been pretty great for humanity it's worked out very well uh so a lovely idea that uh i guess humanity went from being almost a single player experience to being a more communal experience and we can work together more effectively for agents what does that mean if you have more agents coordinating with each other uh what that give us? How do we benefit? Think about anything that we do today as humans, to your own point, Richie. I mean, it's been an awesome 300,000 years, but it's been an even better 70,000 years for us. And the reason for that is because we work really, really well in teams. And think about anything that requires a team of subject matter experts to come together to solve for something. Drug discovery, chip design, incident response, something goes wrong I mean it's something as mundane as that all of these situations I mean no single person is the expert

in all of these situations there are people who bring in expertise like in chip design there are like very long players, there are people who deal with manufacturing, there are people who deal with marketing, there are people who deal with finance and legal and all sorts of functions that exist within the organization. But now, all of these team players, the way they succeed is by doing those three things. They align on an intent. We're going to launch a chip to the market. They align on terminology as part of the intent. They say that this is what things mean. When we say we're going to ship in Q4, it's not that the engineering is ready in Q4, because that's what engineering would say. We're ready in Q4. But guess what? There needs to be a financial plan. There needs to be a legal plan. There needs to be a marketing plan, a launch plan. So all of these, what does readiness mean? Is something that you need to ground yourself behind.

So grounding, coordination, negotiation, resolution of conflicting information, all of these things are things that humans do pretty naturally, but agents don't do today. Agents today can connect together. they cannot think together because they cannot do simple tasks as grounding, resolution, coordination, negotiation and so on and so forth so that's one aspect the other thing is context context matters context is relevant to not just the expertise that a human brings or an agent brings but context matters even to the organization that you're part of the goal that you're trying to achieve so context and coordination and alignment are the things that agents don't have today and that is something that agents will get which allows them to participate as first class citizens in multi-agent human teams which is where the future is going to be the future is going to be multi-agent human teams that are going to solve the really hard problems for humanity

like drug discovery and materials discovery and so on and so forth I feel like that engineering example is the result of like so many misunderstandings around like why things are delayed. It's like, oh, the boss asks, when's this done? Engineering says, oh, yeah, end of Q4. And actually they forget that there's like three months of other things afterwards before it can shoot. But, yeah, I do take your point that like almost anything worthwhile doing, it involves several people at the moment. Well, maybe maybe like hundreds of thousands of people like, you know, people scale very well because we can communicate. and so with agents it's going to be a similar sort of thing. So tell me what needs to happen, like what do you need to put in place? Like you said with humans the cognitive revolution is all about like language to make people be able to communicate better. What do we need to have this large-scale coordination of agents? Agents computing is moving through an arc where we started with isolated agents and this is when agents first came around. But they were pretty siloed and just like humanitarians, they were like individuals working in their well and so we went from isolated agents to

islands of agents to connected agents and now we're looking towards going towards coordinated agents so that is the arc that is forming so isolated agents we just talked about islands of agents people are very familiar with this if you go to any hardness a cloud code or a codex or whatever you can create a whole team of agents that are working together but those agents are an island they are within that harness they don't jump across that harness and talk to other agents that might exist somewhere so the next thing we decided to do as an industry is move towards connecting agents so connectivity is real now so we have projects like agency which we which we created and gave to the learners foundation that's actually spelt A-G-N-T-C-Y for those who are searching. We have projects like NCP that allow connectivity to tools. Projects like A2A that allow connectivity between agents. So all of these projects, all open source, are solving the connectivity problem, which is can I discover agents with

certain expertise? Can I give them identity and access? Can I allow them to message past between themselves? And can I observe them? Are they up? They're doing things or they're not? And so this was step number three, going from islands to connected. But then we are trying to solve the fourth bit here, which is coordinated agents. So coordination is actually taking shape now. So connectivity is real. We've had a bunch of use cases. Coordination is just about taking shape. And we just published a paper which shows throw a bunch of agents at a problem. and 100% of the sessions between these agents have at least one cognitive failure, one coordination failure and this failure rate and the harm it causes actually dramatically increases when you go from two agents to three agents to four agents and so on and even at three

agents what we found was three agents were not aligning on a simple goal more than 16% of the time that's just terrible so the numbers are pretty bad we can connect agents we can make them talk to each other make their message pass but coordination is a real problem and that is what we're trying to tackle and finally once you've coordinated agents the next step in the arc is moving towards expertise which is one thing to say we are all subject matter experts we all coordinated we know where to go but we all live in the real world and in the real world there are two things that we need to be careful about the first thing is regulations and rules so all everything around guardrails and trust and security are paramount for all of these agents to work together I know the rules and regulations of the US I will not do something that goes beyond those or go beyond the rules and regulations of Cisco

as a company and that's one aspect the other aspect is acceleration so I do certain tasks every day in my head when I come across the same task for the fifth time the sixth time I know how to take shortcuts so expertise is not just knowing the subject really well but knowing what's possible and how can accelerate that but also knowing what's not possible and what i shouldn't go and explore so we're calling them cognitive accelerators and guard rates and those things are paramount to making sure that these multi-agent human teams do work in a shared reasoning environment but do so under certain constraints and guidelines okay yeah so i certainly see how does on the one hand you've got all this possibility of like oh we can solve amazingly difficult uh problems once we have like many agents working together but also seems like coordination is like a really tough problem to

crack that's why i like the idea so you mentioned cognitive accelerators like this is what you should do this is the possibilities and uh guardrails like this is what you shouldn't do i'm curious as like what the the structures of command between these things like is there an organizational design thing here like with humans uh often like in companies you have like a hierarchy of like the bosses in charge and uh and commands filter down because of things like a uh collective where it's a bit more i guess if you dip it and everyone votes on on what to do like uh what's the what's the structure of uh the agents then like uh is there a change of command there just like humans and the way you just uh hinted at this we as humans also have different command and control structures in place and there's a spectrum of things here so there are pretty strict top-down mandates that people work with I mean the army is a great example of that where there's a pretty strict command and control and you do not veer off the structure at all on the other hand you have what we are

pretty familiar with also in the tech space communities in open source and and they are pretty much driven by what is the need of the hour and what needs to be solved and people rallying behind that need and building code around that. So that's the other end of the spectrum. And if you really think about enterprises, I mean, people think about enterprises as command and control. It's not really the case. There is a hybrid model here in enterprises where you have certain teams that are somewhat command and control. there's a lot of liberty within those teams to go and color outside those boxes but then take a simple example as walking through a two million dollar deal that flows through a deal desk in an organization in that sort of environment you will see engineering and finance and legal and marketing and all of these various team sales coming together to move that deal through the deal desk towards the two million dollar deal

and there is no real command and control here there is pockets of commands and control especially within similar work but then there's a lot of coordination that is happening within these teams to move towards a common goal so the design patterns for agent deployments are actually not going to be any different from design patterns and human deployments where you'll have strict command and control usually this is within a particular harness you'll have complete openness and this is you've heard things like mold book that came and went but you see a lot of the personal agents move in that direction and then there is the hybrid middle which I think is going to be the biggest piece of the puzzle where you'll have pockets of command and control for similar work and then you'll have coordinated outcomes across teams and across different kinds of work and so those are patterns that we are moving towards and the way we think

about this is the way the internet was formed we don't think of solving these through one giant orchestrator because that's not scalable scalable designs are always decentralized and are always distributed so the way we are thinking about this is through an internet of cognition, which is can we break the problems down into scalable components where there is no single owner, but you actually, in practice, during runtime, move towards eventual consistency, move towards a shared intent, move towards shared context and shared reasoning in a natural way in a distributed way in a decentralized way rather than having a centralized piece of control Okay so it seems like there are a few different models and I guess like the command center thing is maybe the easiest to implement. I guess we already have that with agents. It's like you can get it to spawn sub-agents, whatever, and that's fairly straightforward now. But yet, getting this more distributed system, this seems

like technically a lot harder. This is the kind of stuff you're working on. okay so it sort of sounds like an infrastructure problem getting agents to talk to each other but some of the stuff around designing like what the pattern of agents is and choosing which agent to go where this seems like maybe a high level problem so who needs to be involved in in solving this like which teams or roles need to care about this the great thing about software is that you can codify rules for humans and human teams and human collective intelligence the way we codified collective intelligence is by language through language and making sure that rules and regulations and trips and tricks and tips of the trade are codified in some playbook somewhere but it's an ambiguous problem to solve for humans for agents agents are like humans in many many senses of the word because we just like humans

they converse in natural language just like humans there is no deterministic input or output you can say find me the best flight god knows what is the best flight it's left to interpretation you can also say the agent can come back and say this is the best flight with 90% confidence and yeah, it's the best outcome for the time being but it may be different tomorrow so there's no deterministic input or output so in many ways they're like humans and in many ways just like humans they change shape, they change identities and so there's a lot of infrastructure to be put in place but also they're unlike humans because they run on machines so the way I like to phrase agents is like they are human-like but they run at machine speed and scale And the fact that they are running on machines means you can codify in infrastructure all of these requirements and all of these potential in an infrastructure stack. So what we are trying to do here is through the Internet of Cognition, building out the

layers where you can codify the infrastructure to allow agents to share intent and coordinate, to share contacts and have knowledge and then to share reasoning and move towards collective intelligence so we are building that entire stack out and the way we are building that out is just like humans we are first building the language for agents and you and I know this but the language for machines is protocols so there is a whole set of protocols that we are building out we are calling them layer 9 protocols if you are a networking geek like I am, you know that networks so far have been built to carry data. Networks were never built to connect intelligence and carry cognition. They were never built for that. So we are adding a few layers of protocols that allow networks to evolve towards connecting intelligence endpoints

and carrying cognition state across those endpoints. and these are protocols that allow you to coordinate, allow you to get grounded on terminology, allow you to negotiate and do all of those things that humans do. So that's the protocol layer and we've just announced a couple of protocols. One is called Accord, the other is called Concord. I'm not going to get into details but these are protocols that allow agents to do that. The next layer up, what we're trying to do is build a fabric out. So like I said, scale at this layer is going to come from distributed systems principles It's going to come from decentralization So the way to scale this would be to build a fabric Where humans and agents can talk to each other using those protocols And then also attach to a few things Attached to memory, where you can deal with shared knowledge and shared context And then attached to COGS and GATS these are cognitive accelerators and guardian technologies and I'm borrowing the terminology

from the Turing Award winner Raj Reddy who said AI should be just of two types either a guardian angel or a cognitive accelerator and we completely agree with him so we're building those things out where you attach to this fabric memory and you attach cords and guardrails, GATs, so that you innovate and reason within guardrails and within what's allowed and what's not allowed within the organization. So this is a stack that we're building. And we're starting with the protocols. We're going towards the fabric. And then we are going towards Codz and GATs, which we call Cognitive Engines. Okay, so it sounds like historically there have been a few layers missing. So if you wanted to get agents to talk to each other, you've got to get sort of very deep in the infrastructure weeds and this is just letting you get a bit closer, at least talking in the kind of business language you want. So this is how I want to solve the problem. The cognitive acceleration is like, this is what I want to do. The Guardian Angels, this is what I don't want to do.

And then the memory is like, I guess, this is a background context. That is like, I guess, like product requirements type stuff. And these are all the facts of what you have to do. Is that about right? That's about right. And the context will also evolve over time. So the fact that you have agents and humans talking to each other over these protocols through a fabric and you attach memory to it also allows you to capture emergent context. So the one thing that we want to do is not just limit ourselves to the knowledge and the memories of these individual agents, but we want to capture the knowledge and wisdom and even the episodic memories of the entire multi-agent human team as an emergent system. because that's where the collective power is. So these memory elements that are attached to the fabric learn over time so that they are building context over time as an emerging system rather than a point-in-time view of the world or a subject matter expert view of the world,

which is how human teams typically operate as well. Okay, yeah, so this is not just a way of having agents communicate with each other. It's going to be having humans communicate in that as well as you've got this mixed human and agent team. Okay, I know you're still in the process of building this stack, but do you have any examples of real-world use cases that look like how you've tried using this system already? We've been working on this for a while now. We started working with the connectivity layer, which I talked about earlier. The open source project is called Agency. You can go to agency.org, check it out, and please contribute and give feedback. But that connectivity layer allowed us to deploy real multi-agent systems in customer environments. And when we started deploying some of these multi-agent systems, we solved problems in the cloud native space, we solved problems in the networking space, in the healthcare space, in the video surveillance space.

And all of these use cases you can find on our website. But what we quickly realized is this gap that I talked about earlier, which is you can connect agents but you cannot the agents cannot think together so they can talk to each other but they cannot think together and so what we've been doing is running some experiments to figure out where the gaps are and some of the stats that i mentioned earlier point to those gaps the fact that even though all of the individual agents in your multi-agent system can absolutely be right and correct in what they give out as an answer if they are not coordinated if they are not aligned on intent if they are not verifiable as a system as an emergent system you don't gain the benefits of a multi-agent system and in fact what we see is even though each individual agent is actually giving you the correct answer the fact that you are not aligned reduces the efficacy of that multi-agent system to less than 16%.

And so agents don't have to be wrong, they just have to be misaligned for you to just suffer and just go through this problem space where they're not working together. So what we're trying to do is we're building the stack out and what we've released so far are the protocols that enable you to talk to each other at the semantic layer. and we've looked at the protocols and we said there are many ways to deal with this the first and foremost way to enable protocols is to make it natural language based because that is the lowest common denominator you and I use natural language agents talk to us using natural language so the lowest common denominator is the protocol layer that will work the best and the widest will be based on natural language so both of these protocols deal with natural language processing to enable coordination between agents what we realize though is that that is just the beginning because like I said earlier

agents are like humans but they're machine code so we can do better we don't have to go towards the tokenization cost of moving from latent space to tokens that humans can understand you need to do that only when humans get involved But if agents talk to each other You can actually build protocols That are agent to agent protocols At the cognition layer That do not go through the Organization packs And so we built a whole bunch of protocols Called latent space protocols That actually They are similar to you and I having A neural link implanted in our brains Where we don't have to convert ourselves Our communication into English But we can just think and you'll get everything that I'm thinking through and that's much easier to do with agents because they're software-core running on machines so we've introduced a few things we are going to introduce the fabric to scale out a little bit later

and we are starting to introduce some GATs and COGs as well so we've introduced a GAT that gives identity and access based on semantic communication and this is an interesting one, Ritchie where if you think about identity and access you and I get identity and access that is long lived Cisco knows who VJOY is and will give me identity and access based on VJOY for the duration of my employment here agents are different they'll come in they'll take on a certain shape and form they'll take on a different shape and form the next hour, the next minute, the next day and so the way we are giving identity and access to agents is based on the task that the agent is trying to achieve so today it might be just reading my credit card bill tomorrow it might be actually trying to use my credit card to make a transaction

but the end point could be my agent on my laptop and the backend running in cloud somewhere those endpoints are not changing what's changing is the semantic communication taking place between the mobile device and the backend in the cloud and at one point it's saying let me read your card card history the next minute it might say I'm going to use your card card to do a transaction so can we give identity and access or access controls based on the semantic communication that's happening between these endpoints and not just the static endpoints themselves. And so we're calling this CASA, which is continuous, and I'm going to butcher this acronym, but it is basically doing continuous semantic authorization and giving access based on that. And so there are some of these guardrails and GATs that we're also releasing, but you will see an entire roadmap from us

in the protocol space, in the scale-out space, and in the COGS and GAC space over the course of the next year. Okay, so I have to say, I love your idea of like the latent space communication being like a neural link. I feel like that'd be a great way to record podcast episodes if we have a link between us. Maybe we have to translate it back to English for the audience afterwards, but it just seemed like a fun thing there. Although, so you mentioned around the security around these things, so you give permissions based on tasks. beyond this sort of coordination of agents stuff, is that a sensible thing to have like whenever you're using agents that you should provide task-based limits on what the agent can do? A thousand percent. I mean, one of the things that we're going through as an industry is there's a certain bit of a split between what agents can do and what they're capable of versus the amount of trust and confidence individual and enterprises can place on those agents. and sometimes there is this bifurcation in the path

where you say let me go with agent capability sacrificing trust versus let me go with trusted agents even though I might have to sacrifice some capability and that's where we stand today but that's not how it should be you can have it all, you can have trusted agents that provide the maximal capability from the best frontier model out there and you should not have to choose and the reason why we are at this predicament today in the industry is because of a lack of some of these things in the infrastructure the fact that agents cannot talk to each other or humans and agents can talk to each other but it's it takes a lot of time to align behind intent it takes a lot of time to really figure out what I am asking the agent to do even personal agents what the agent is actually trying to do there's a lot of back and forth that happens and then once the agent is actually

in a loop trying to do some of these things I know we've all heard the term human in the loop to human on the loop and so we all want to get towards human on the loop and we want to reduce the number of interrupts the human on the loop has to make when those loops actually deliver value but we are not there yet and I think there are aspects of this loop that can be enabled within islands but the world is not full of islands. I mean the example that I just used, you'll have multi-functional groups of human agents coming together to solve for a business need. You don't just have developers, you don't just have legal folks, you have developers and legal folks there are different terminologies and different guardrails and different accelerators that need to come together to solve for an outcome and so in that world you will have to make sure that this infrastructure is built out properly and identity and access is the

first problem the first hurdle that enterprises run into and so that's why for us CASA which continuous semantic authorization in access I don't know which order it goes in which but CASA is so critical for us because it removes the first hurdle it actually looks at semantic communication figures out exactly not just the task but the tool that the agent is trying to access but also the transaction that taking shame because an agent might be just one entity in a multi human chain that is actually transacting something in the real world. So I might not be the originator nor the end point for that transaction or for the task or the tool. I might be just an entity in the chain of transactions that might be taking shape. So we have this notion of T-back, tool, transaction and task-based access control

and this is exactly what it's doing. It's parsing semantic communication and it's either raising authority or bringing down authority based on the task, the tool or the transaction that's taking shape through that agent and that's just the first hurdle there are many many many security and governance hurdles that we need to cross as an industry and we'll be building those gaps out but also let's not ignore the accelerators because the other flip side of this coin like I said is the accelerators and today when agents come at a certain task they restart from scratch every single time and what we need to enable is an infrastructure framework where you build that body of knowledge you know what needs to be done you automatically pull certain things we know what HIPAA compliance means we don't have to reinvent HIPAA compliance from the ground up you know exactly

what needs to be done humans have done this for ages can we accelerate agentic outcomes in areas and in situations and verticals where HIPAA compliance is key without building that stuff from the ground up. So accelerators as well as GATs are going to be key in making sure that multi-agent systems work and they quickly move towards working together with humans to drive and deliver on outcomes that matter to humanity as either humanity-based outcomes or enterprise-based outcomes. Okay, so I have to say, from previous episodes of the show and from talking to AT&T customers, perhaps the biggest blocker to agent adoption in companies is like fear that agents are going to do something stupid, they're going to cause a problem. And it sounds like from what you're saying, security and like a lot of the guardrails around agents, it's at least theoretically solvable. Like how close do you think we are to having like this not be a problem anymore?

One of the things that we've been looking at through this whole Internet of Confirmation work is again how do we make the non-deterministic parts of a multi-agent system more deterministic and I know it sounds controversial and contradictory but that's where you want to be where the confidence level goes up in determining what a task is what a transaction is, what tools are being accessed so that then you can do not just identity and access but even simple things like segmentation and making sure that the right topics are being discussed with the right teams. And the teams, when I say teams, it's either humans or agents or a combination thereof. So simple things like, Richie, you and I can talk about public phasing information on Cisco at any given time but I cannot discuss confidential Cisco information in this podcast. But today's networks and today's computer infrastructure, in fact,

today's cloud infrastructure actually are not that fine-grained. And so a lot of the fear that comes out, a lot of the paranoia is because of these reasons where we've dealt with endpoints, compute endpoints, as data endpoints. We have not dealt with endpoints as intelligence endpoints and that foundational jump is something that all of infrastructure has to make. Well, what does it mean to connect or to enable intelligence endpoints? It means treating them as ephemeral entities. It means treating them as semantic endpoints. It means parsing the intent behind what they are trying to achieve and then enabling your firewalls, enabling your hackles, enabling your identity and access, enabling your resource distributions based on what the semantic intent of these intelligence

endpoints happens to be. And so in the previous example a simple segmentation which says Richie you and I, VJOY, are connected and in this session the fact that we are connected we can talk about anything is what we would do in a data and point connectivity scenario but really what we want to get into is anytime i start discussing confidential cisco information something should pop up and say it's not allowed in fact it's the simple way to think about it is like a bleep and something should come in and say bleep i bleep that out because that is a piece of information that we should not be sharing in this communication channel even though the entirety of the infrastructure that today sits between you and I Ritchie allows me to say something confidential today in this communication channel but we cannot be there because agents a Ritchie agent and a VJOY agent are not fine-tuned to that today so the moment a VJOY agent starts talking confidential

infrastructure should step in and bleep that communication out and that is the infrastructure that we are building and it is codifiable its infrastructure its protocols its its memory its context its fabrics its its cognition engines this is all software we've built software in the past we've been building it for several years we can codify all of this in the infrastructure so that agents can go wild but the infrastructure actually protects them from going wild and so it provides them the swim lanes to then do their best thinking within the gardens and the swim lanes that they can play okay so i yeah i say i would love for you to give me some juicy confidential information about uh but yeah i suspect what will happen is if you do that i will get a very strongly word of email from someone the Cisco comes in yeah so it sounds like what you're saying is the

security properties, the limitations shouldn't be a property of the agents themselves they need to exist at a lower level in the infrastructure so it's like whatever the agent does there is like a sort of governing body underneath it that says okay no you can't do that. I think the security is the first hurdle because that is paranoia and rightfully so because I don't want identity and access and I don't want communication that shouldn't go out is going out to a different agent that shouldn't be going out but that's the first hurdle I think the next hurdle that you'll come across is great, I've connected agents they are communicating semantically through IOC, hopefully and then you also have security guardrails and you have accelerators in place but then you have to go to the next step which is, are these multi-agent systems doing what they're supposed to do? Is the after picture really better than the before picture? Can you verify whether they are completing the task?

Can you verify that they are aligned on intent? Can you verify that they are all grounded? So that's the next step. So, and that's also a semantic problem. Because verifying an outcome implies verifying the outcome from every individual agent, but also measuring it against the shared intent measuring it against the coordination tax that needs to happen measuring it against the negotiations that are taking place one simple example I have today is this whole observability and verification state of the art today for some reason people are after tokenomics what's the tokenomics of this situation and we should be spending less and getting more out of our agents And I agree, tokenomics is a big deal. I shouldn't be spending more than I should for a certain outcome. But tokenomics is about faster and cheaper. What about better? You are missing the better part which is where you want to be.

So yes, cheaper, yes faster. But I want better as well. Faster, cheaper, I am better. And so tokenomics can drive the cost down, make things faster. But to measure whether things are better, the after picture, you need a semantic understanding of the world. So that's the next step and so on and so forth. And finally I would say the final step would be emergent behaviors and this is where I talked about having memory and knowledge and wisdom that is captured for the entirety of the multi-agent human team rather than being limited to a singular agent or a singular human and this is where that is the holy grail in my mind the way to get to super intelligence is not just scaling individual intelligence but scaling intelligence in a distributed way and the only way you can scale intelligence in a distributed fashion is by extracting

emergent knowledge and emergent behavior and shared reasoning across a group of multi-agent human systems. To me that is the third and final step. So getting to security and safety and guardrails and governance because that will get you inside the door. Then getting to better. What is better? Is it truly better at the semantic and cognition layer not just at the observability layer? That is the efficiency knob. And then finally, once you're comfortable with security and safety and efficiency, then getting to emergent behavior. Because that is really where you should be playing. Where you should be taking a multi-agent human system and solving for problems that each of us individually cannot solve. It's a very compelling vision. I mean, of course, like the intelligence of the whole human race is much more than like any individual human. And so I love the idea that once we start being able to have this coordination of scale, then we're going to have the potential to solve amazing things that we can't do now.

So we talked a lot about the technology side of things and some of the structures. One thing that has cropped up a few times is you've been talking about how we're going to have teams of humans and agents working together. Do you think this is going to be a fundamental shift in what a team consists of? One of the things that we've been seeing and we're either painfully aware or excitedly aware about is the fact that agentic computing is reducing the time it takes to produce artifacts. and so whether the artifact happens to be code whether the artifact happens to be a legal document a marketing brochure, artifact production time and cost and energy is actually from the human end is actually tending towards zero and that jump has taken shape it's been like I would be surprised if there's anybody out there that is not using agents

to produce artifacts today and if you are one of those I would say catch up the train has left the station I need to run and get on the train but if you think about an organization and the productivity towards an outcome that an organization needs to move towards there are three pillars that are needed to go towards that outcome the pillars are the unit of productivity productivity, the substrate for coordination and the hierarchy or the design pattern of the organization that we touched upon a while ago. So there is a productivity unit which used to be really large teams and for hard engineering those teams are still pretty large. If you are sending a rocket into space, you cannot do this with small teams. You need all kinds of experts all coming together, working together

to send a rocket in space. So hard engineering, building a car, building a router or a switch, building a chip like a GPU, I mean all of these problems are hard engineering problems and you still need large teams. But for software, that productivity unit actually already shrunk a while ago, a couple of decades ago in fact, where we moved towards two pizza teams. And during the cloud service and the cloud software era, there was no longer a 30-40% team delivering on cloud software. We had team sizes of 8 to 12 people and these were the two pizza teams and this was pushed by Amazon and rightfully so. and we are all aligned behind two pizza teams before agentic came into play and now with agents that team size especially for software not for hard engineering but especially with software that team size is actually shrinking even further

and there's an extreme viewpoint which says you should go with a team of one we believe that a tiny team with token as we're calling a T3 tiny team with tokens is the right size so what is that size? the size really doesn't matter but it's roughly between 1 to 5 humans augmented with multi-agent systems and why is it 1 to 5 humans? because you are actually biasing on what is truly hard even now and that is judgment, taste, influence, accountability so you bias on that and you decide based on that whether this T3 requires just one human and a bunch of agents or it requires maybe three to four humans and a bunch of agents. And so for us, the productivity unit has shrunk toward T3, which is roughly between one to five

humans augmented by a whole bunch of agents and tokens. So that's the productivity unit. but what's happening Richie is since the productivity unit is shrinking the coordination surface and the coordination tax is exploding and nobody is talking about the second part because if you think about coordination there are two parts of coordination intra-team and inter-team yes smaller teams means intra-team coordination has is tending towards zero but the inter-team coordination has just exponentially exploded because now we have an end squared problem and so yes you have many many small teams the smaller the team the worse the problem but you need to coordinate across all of these teams to drive towards a business outcome because one of the things we learned the hard way so we introduced T3s in our organization

and singular team has a canary worked really well you will see teams using T3s as a productivity unit Amazing The moment you start 10 teams and 20 teams As T3 sizes The coordination packs Making sure that you are all aligned And working towards a common deliverable and common outcome That just went through the roof And so what we've done is we've created this whole T3 operating model which solves all of these three pillars shrinking the team size to a T3 making sure that there is a context layer that allows for context to be maintained and then putting a coordination structure in place where we have context leaders that maintain consistency that maintain shared intent that maintain shared reasoning and don't allow people to just team these T3s to just go haywire

was in the end you're shipping a business outcome. We built that in our organization as a structure. Sounds familiar? Because all of these three things are also problems that we're solving in the internet of cognition for the aesthetic world, but we are also solving it for the organization for us to then be able to coordinate and move towards common outcomes. Yeah, so I think I'm okay with smaller teams as long as I still get to expense like two pizzas. that's fine you're too the puzzle might be going towards tokens but that's maybe maybe so yeah I like the idea of having smaller teams and I agree that it can be very difficult communicating between teams now I was thinking about like when I try and promote these podcast episodes like try and coordinate with other teams now rather than speaking to the marketing team I've got to speak to like product marketing and life cycle marketing and paid marketing and growth marketing like there's about five different teams so So, yeah, does this work, this idea that you have smaller software teams,

does that work in the rest of your organization? Is that going to work for your marketing, sales, finance, all the rest of it as well? So all of us have actually moved towards T3s at this point in time. And we started working on this, I won't call it a reorganization, it's just an operating model. So the teams still are functionally organized, like in any other enterprise. but the way we solve these three pillars and make them work together is through an operating model and what we've said is E3s can be formed by anybody and this gives you the liberty to go and experiment because one of the things I'm a big believer in is AI and agents don't just allow you to do your function faster and more efficiently but they actually allow you to expand what you do. So a software developer is not just producing code faster and better

through agents and models but a software developer can actually do a little bit of product marketing can do a little bit of product management can go and talk to customers and do customer feedback they can actually expand their capabilities and do a lot more than they could in the past and the same thing applies for other functions so a product manager can actually produce code can produce designs can actually take those designs to code to a customer and do the loop completely on their own so it is expanding the capabilities of every individual of every function in the team so what we said is that a T3 can be formed by a product manager, a marketing person a software engineer or a combination thereof. We don't restrict that. So T3s could be a mix or could be just one function. Doesn't matter. Now context. To make sure that we

enable context sharing, we built a context layer. We are calling it TOM. It stands for the organizational memory engine. And the reason we picked this is yes, it is the organizational of memory engine but it is the tome because it actually captures everything one of the learnings we've had is today the way we work with collaboration surfaces is teams use emails slack or teams or WebEx in our case they use the entirety of the Atlassian suite like Jira and Confluence they might be using SharePoint or Google Docs and there are many many many collaboration surfaces that we all work with and so when we started building Tome we said human teams agent teams T3s all of these collaboration surfaces and then Tome as

a central repository of everything that is needed for us to collaborate what we realized is that not all collaboration services are built equal and just like hallway chatter is the freshest signal you can get in an organization the equivalent of that in this hybrid virtual world that we live in is slack or teams chatter or the video meetings that we all do all day long so we prioritize There is a collaboration hierarchy, an authority that we've built into Tome, which says, here's a freshness score, where Teams Chatter and Video Conferencing Chatter is the highest when it comes to freshness. and then there are certain documents that have authority score which if it is being created if a document is being created or modified by a context leader and we've identified context leaders

if it's been created by context leaders they have the highest authority score so there is a freshness authority or a hierarchy and there's an authority hierarchy and that's how we are dealing with this hybrid mode where we have T3s which are humans and agents this massive collaboration surface that exists today and we have to work with and then TONE that is the wherewithal and the be all and end all of all context in the organization the challenge that I've given the team is in a year's time I want to see the model flip where you have human agent teams, the T3s, write to Tome and only Tome and then Tome actually goes and writes to a PowerPoint or to a Word document or a Google Docs or an email or whatever it be only if external consumption is needed.

So flip the model where Tome is your first context layer and you only use some of these other collaboration tools when you want to communicate with the outside world in some way shape or form. So that is something that we are building today and a bunch of learnings there and the third pillar which is the coordination layer the way we solve it and I just hinted on that is we've identified context leaders and the way we've done that is that T3s every organization works on some form of objective and key results OKRs so the objective leaders and they are area leaders and then they are T3s so area leaders and objective leaders their entire job is to ensure that things are streamlined when it comes to intent and things are streamlined when it comes to the KPIs and the goals that that objective is trying to produce. So by doing this, we are solving all of those three pillars and we can run

fast as well as be coordinated in attempts and running fast. Now, this has been only eight months in the making. We've already learned a bunch of lessons but we've rolled it completely in our organization and we're seeing benefits of that even today. This is a very, very cool project. It sounds like a fairly substantial project as well, like building this sort of, I guess, a context layer for the whole organization. So is it basically, is it like a fancy data warehouse? Or what's it consist of? Think of it as memory layer that you would use today in agents or you would use today within a harness. So it consists of a layer that evolves over time, just like markdown files and vector databases is evolve over time in personal agents or in coding agents for example but the key thing here is A there's connectivity to everything in every collaboration surface that we use

B it works really well with the context leadership that we've defined so that it does not hallucinate it does not just summarize but it knows freshness hierarchy and it knows authority hierarchy those things are key and the third thing is over time as we pivot to more and more agent forward approaches this is going to be the context layer for the multi-agent system the So it is completely moving towards an agent native way of working rather than today if you think about all of the memory layers that exist, it's still very human driven and the way we're building this is it's agent forward rather than human forward and so that's where we will get towards but right now it's still very very human driven and the way we're thinking about this is over time this is going to be the emergent knowledge of outshift

moving forward it's not going to be the knowledge contained within Vijoy's harness or Richie's harness or somebody else's harness this is outshift's harness over time so throughout this episode we've been talking about like a pretty big idea which is like how you move towards having like large amounts of agents working alongside your employees and all coordinating together. For anyone who's interested in this, just to wrap up, how do you get started? I would say the first thing you should do is head over to our website, outshift.com, because we are an emerging tech organization. We are the incubator for Cisco, looking at emerging technologies and figuring out how and where should Cisco play in these emerging technology spaces and what kinds of businesses Cisco should build in these spaces so we are constantly dealing with technology risk

tech that is risky but how do we make it less risky cognition is one such area quantum is the other such area and we are also playing with market risk where market forces are shifting the market might not be stable it's evolving pretty rapidly it's not stable enough to then take big bets so we are here to reduce that risk to figure out what are the big bets that Cisco should take. Given that charter, a lot of what we do is education. Given that charter, a lot of what we do is experimentation and pointing the spotlight in areas that are interesting but also balancing that with the reality of the situation. So go to www.outtrip.com. That's your first stop shop. in this space I would say go to agency.org we've talked about that it's agntcy.org that's the open source manifestation of a lot of the work that we're doing

so join us, give feedback, use it that's the other place I would say the third place is I write a lot about this in my sub stack as coherent cognition I write a lot about this on LinkedIn follow Outshift, follow me you'll hear our journey in a very non-pitch manner because we are not selling anything. We are selling the technology. And we are all excited. Bunch of technologists. We want these things to work and we want all of us, your listeners, to come in and help us make this happen. So that's where I would go. Okay. Great question for our listeners. That is, yeah, go and try the technology and see if you can make use of it. Build something cool. All right, wonderful. Just finally, I always want more people to learn from. So who else's work are you excited about at the moment? The one person that has really shaped my thinking when it comes to Internet of Cognition

is Michael Tomasello. And he's not a technologist, but he's looked at the evolution of human intelligence and how we've gone from individually being intelligent to collective intelligence. and that arc that I just talked about at the beginning of the show quite a bit of that is shaped through his thinking as well so I would say look at the work that Michael Thomas Seller has done I would say Michael Levin is the other person who looks at a whole bunch of scenarios where intelligence is an emergent property where he says if you think about this is an extreme example where he says if you think about cells in our bodies they are individually intelligent but they do not do what VJ can do and so there are pockets of intelligence in the cell and believe me when you listen to him there's a podcast that he does with Lex Friedman

I suggest everyone here go and listen to that podcast because he talks about the intelligence that actually lives within cells as an organism and it's actually amazing cells are not like a piece of stone they're actually intelligent but together they build up Richie and they build up Rejoie and they build up all of our humans and they do exponentially larger better things than what they can do individually so I think Michael Levin is the other person that I am a big fan of between Tomasello and Levin those are my two big influences when we thought about the integrative cooperation it's great to kind of break out of the technology bubble sometimes and just think about some other scientific ideas of course yeah as we're thinking about like what happens with artificial intelligence is that we're thinking about other kinds of intelligence beyond just like modern day humans so yeah I love those ideas

for I would say I'm not the only one there are a whole bunch of computer scientists academics as well as Frontier Lab folks that have actually looked at studies like this, like Tomasello is actually pretty widely cited if you start looking at some of the intelligence and cognition work and super intelligence work that's happening in our community so we all draw from certain sources of inspiration but to me the draw has been primarily around the pursuit of super intelligence being a distributed and a collective endeavour and not an individual endeavour. Thanks for a great way to wrap things up. Wonderful. Alright, thank you, VJO, for your time. Thank you, Richie. This has been an amazing conversation.

you