From Multi Agent Systems to Institutional Learning in the Enterprise - with Papi Menon of Outshift by Cisco
The AI in Business Podcast · 19 Mar 2026
Source: https://podcast.emerj.com/from-multi-agent-systems-to-institutional-learning-in-the-enterprise-with-papi-menon-of-outshift-by-cisco
Transcript: https://podcasts.opennash.com/shows/ai-in-business/from-multi-agent-systems-to-institutional-learning-in-the-enterprise-with-papi-m-ad4966/ (via OpenNash Podcasts, machine-generated)

Welcome everyone to the Emerj AI in Business podcast. Today's guest is Papi Menon, Vice President of Product Management and Chief Product Officer at Outshift by Cisco. Outshift operates as Cisco's internal incubator for emerging technologies, focused on reducing risk and delivering innovative results. Papi joins Emerj's Matthew Demello to explain why enterprises see early wins with agents, but struggle to scale them. Today, systems can connect, but they can't share context or learn as a coordinated whole. He lays out what it will take to move from isolated agent performance to collective intelligence.

He also shows how leaders can make progress now by choosing low-risk, high-impact starting points and building an open, interoperable foundation that won't trap them as the technology evolves. Today's episode is sponsored by Outshift by Cisco. For our AI solutions partners, position your brand alongside the Fortune 500 leaders defining the enterprise AI roadmap. For the opportunity to showcase your solution to the executives currently funding and scaling global initiatives, partner with Emerj. Secure your partnership at go dot merge dot com slash partner.

That's G O dot e m e r j dot com slash p a r t n e r. Now, the conversation with Papi. Papi, welcome to the program. It's a great pleasure having you.

Thank you for having me, Matthew.

Absolutely. In our first episode in this series, we explored why enterprises are moving from that that single model system to multi-agent systems and where early value is starting to appear, where infrastructure challenges like like discovery, identity, access, and observability uh begin to surface. For today's discussion, we hear from so many leaders on the show who feel that they're doing the quote unquote right things, right? Pilots succeed, agents perform well, and and teams see localized wins, despite the bad news from MIT and all these places saying 95% of pilots uh uh fail and or die on the vine, yet progress stalls and organizations find themselves repeating work rather than building momentum.

And we're seeing this in a lot of different places. Maybe that's a little feeding a little bit into that ninety-five number, you know, just from your vantage point, seeing this from Outshift by Cisco, your colleagues at Cisco. In what ways are we seeing enterprise AI momentum slow after early agentic success, and why?

I wouldn't say that the momentum is slowing. It's just that as enterprises try to scale up their agentic deployment, they run into certain problems that need to be solved. And this is something that we've been evangelizing for quite a while now. And if you think about it, it's quite a natural natural progression. As enterprises started to deploy agentic AI, they started to do it with specific tractable problems in different parts of the enterprise. And that was fine.

And and and many of our customers and partners are starting to see some significant success in that. But as soon as you start to go beyond that scale where you have what we call multi-agentic systems, right, where you have complex agents that are trained in specific domains, that are trained on specific data sets, have specific expertise, now starting to cross their organizational boundaries, now starting to cross, talk to agents in other parts of their enterprise, or even perhaps agents that are outside their enterprise.

As soon as these types of data transfers start to happen, the complexity of the problem, just, you know, balloon. And this is something that we've talked about for a while. You know, we've talked about it pithily as kind of the Internet of Agents. How do we enable agents to discover each other, communicate with each other, have all the security and observability that is that is needed when you have? And again, you know, we need to understand part of the disconnect in this whole agentic AI space and in the AI space more broadly has been kind of the magical almost, you know, things that people can do in the consumer space.

I mean, you can go into chat GPT or Claude and and ask it to do all kinds of things. But why aren't we seeing that same kind of progress on the on the enterprise side? There's been a growing recognition and realization that, you know, on in the enterprise, it's gonna be, you know, as Andres uh, Carper said, you know, it's not gonna be the year of the um, the agent, it's gonna be the decade, right? And there's a recognition that it's gonna take a little longer than we thought.

And you know, that's for good reason. You don't want enterprises to be playing fast and loose with their customer data. You don't want them to be, you know, ignoring uh fundamental security principles. We all we we don't want our banks to be doing that or our hospitals to be doing that. We want enterprises to move with with, uh, intentionally and with deliberation, with the right security and observability and all of that in there, and with the right, you know, enterprise-grade tooling in place.

So I wouldn't say that the momentum is slowing. It's definitely a realization creeping in that these things will, you know, once you start scaling it out beyond single agent systems, once you start multiple, you know, having multiple agents in complex multi-agentic systems talk to each other and doing that with all of the enterprise-grade capabilities that we need, then you are gonna have need a little bit more time to get it right.

Yeah, it it might I I I'm with you there in that it might not be slowing down so much as a a reflection of not just misaligned expectations, but also misaligned education of how these systems really work and putting expectations in the right place so that these systems are more in line with their actual capabilities. Rather than, as you mentioned, uh, sort of the magic effect. You also, uh, mentioned, uh, in the last answer a relatively relatively robust and concrete word that we know from neurology, which is like cognition.

And, uh, I I want to put a a more definite discussion of that term. I want to be able to put a a more concrete definition of that term, just because I I think there's kind of a penchant, especially the less you might know about these technologies, to anthropomorphize them or see them as as magic. When actually, when you understand how they work, there are very specific bounds by which these capabilities are disseminating through the e enterprise.

And it makes a lot more sense from the top down once you understand how they really work, that they're not so necessarily magical anymore. You you were also mentioning that coordination. Where are we seeing that breakdown when enterprises are moving from merely connecting systems with these tools to that cognition becoming like something of a central brain of the enterprise?

Yeah, that's a super interesting question, and this is something that we've been talking about a lot and we've been discussing how we take. So we've made a great deal of progress over the last year and a half, two years in enabling agents to discover and connect with each other in a secure, scalable manner. And so we we call that the syntactic communication, right? How do you enable agents to talk to each other in an open, interoperable manner and and do so in a way that they can at least syntactically understand?

You know, when they say certain words, you know, you know, does the noun come first or the verb come first? You know, how do you make syntactic sense of the of what you're talking about? And that's great, and that's by and large working today with the protocols and and tools that we have in place today. The next level of that is what you alluded to cognition, you know, and how how is that the syntactic, you know, the words that you exchange?

How do you now attach meaning and context to that? How you share the contextual understanding of what has happened and what you're trying to communicate. And this is the way that multi-agentic systems can have a shared understanding of what is the mission, what are they getting together to do, and how do they progressively refine and build on that understanding. And so this is that cognitive layer that we are talking about, and it is still very much, uh, something that doesn't exist today, you know?

So today you have a lot of agents. Most agents today are are continuing to improve the way they do things as the models improve, as the their tooling and their ability to talk to other services improves. They can, you know, the the agentic capabilities are improving within their silos. But how multi-agentic systems, how multiple agents can come together and improve as a collective through a shared understanding of what is going on in service of the mission that they've come together to accomplish.

That layer, that is what we call cognition. How do you tie that shared understanding to the mission? And that is doesn't exist today, and that's what we're working on. That's what we we refer to when we talk about agentic cognition.

So j so just in terms of that transition from connection to cognition, now that we have uh a more robust def definition of those terms outside of what cognition literally means, especially to neurobiologists or or anything of that nature. Would you think of an example, maybe of uh, a company that successfully made that transition in your experience?

Yeah, uh, let me address that first part, uh, part of of your question, right? Like, if you think about cognition, the dictionary definition of cognition is really about acquiring knowledge and understanding through the experience of doing something, right? And that that's cognition. And agents today, in in a very siloed sense, have that today, right? They can the agents do have a history, they do learn from their past experiences and so on. But what is lacking is when you have a complex, multi-agentic systems where multiple agents are collaborating on a business problem or a scientific problem, then collectively they are not learning together.

They don't have cognition in that sense. They don't, as a system, learn from each other in a shared contextual understanding of what is going on. And that is, I think, where the the thing that we are, uh, when we talk about the Internet of cognition and so on, that is the problem or that is the opportunity that we are trying to address. In terms of companies that have done this well, I can talk about design partner of ours, uh, I won't name uh names, but this is a major major company, uh, in in the financial services.

They have been doing a network debugging tool with us, they've built it in partnership with Outshift. And the goal here is to be able to to simulate network changes before they are actually implemented. So one of the big sources of network outages is badly configured network changes. So you roll out a change without fully understanding all of the different ramifications of it, you roll it out and then things happen, you know, bad things happen. The idea here is that you can simulate what your network is like digitally, you can create um, a a digital twin, as it were.

You can roll out your changes and simulate it on that digital twin before you roll it out in your actual network. This is a big, complex, multi-agentic system that we have built out with them. It is running in production today, and they've already, uh, realized great benefits from it. What is happening here is that you have multiple agents that are all talking to each other, they're all debugging these, uh, problems, different aspects of it, and they are sharing context, but they're doing it one off in conversations with each other.

What we are now trying to get to with the Internet of cognition is where they can do this in a shared common protocol across a shared fabric so that the entire complex multi-agentic system can learn together as a collective and improve the outcomes and do it much much more efficiently and quickly.

Absolutely, and I think we're still very early days of not just the agentic hype cycle, but also agentic adoption writ large for where it can go just across the global economy. And I think the the messaging on here can really vary depending on where you are in industries or or that that larger macro-scale global economy. Just as just as an example, we had, uh, one of our friends of the show, David Glick of Walmart, come on and talk about how they're dispersing literally thousands upon thousands of agents.

Uh, they call them nano agents for specific tasks. And I think this might lead to a possible fallacy where business leaders think, oh, I I'm just going to be swimming in these things. I'm going to be having more of these than there are employees. And that might not be the case for achieving the business goal. It might not be the case, depending on the kind of system and the results that you want. Wondering how you talk to your partners about kind of gauging the scale here in terms of what's needed for business problems.

And and just really, moreover, why doesn't adding more agents to a problem really solve it? Uh, depending on the problem and even what we're talking about here between coordination, connection, and cognition.

There there's so many, so many different angles in that question. So let me let me tease that apart and address them one by one, right? Like, totally, totally fair to say that enterprises are at varying, you know, at at different stages in the agentic uh adoption, um, you know, journey. Some of them are still very early, some of them are, you know, very early and haven't even, you know, are still dipping their toes into getting single agent systems working.

Others have complex, uh, like ourselves, you know, we ourselves have built complex multi-agentic systems. So if people say, oh, you know, multi-agentic systems are not real, nobody's really running them in production, I beg to differ. I can show you evidence that that is not true because we have done it ourselves. But but having said that, you know, so people are at different stages in their journey. But more importantly, I think to the core point that you were making, that journey will lead to different destinations depending on the shape of the business that you're in and the kind of outcomes that you're driving towards.

So, for your, and we don't try to get into the specifics when we talk to our customers and our design partners. I'm not trying to become an expert in their business, right? Like, that is not the goal here. You, you know, they know their business far better than than I do. What I am trying to do is arm them with the best possible tools that gives them the optionality to solve this problem in the way that they see best, right?

So what we do see happening is that it is going to be a heterogeneous environment. So it is not going to be a case that one model or one way of doing things or one type of agent is best suited for all business problems. It is going to be a multitude of different models, large language models, small language models, sometimes running on your own infrastructure, sometimes running in the cloud as a service, sometimes running in a VPC, sometimes exposed as an API, sometimes through a native agentic interface, what have you?

We are going to see a very broad, heterogeneous flavor of, you know, different types of agentic deployments, all trying to work together to solve these problems. And our goal is to make sure that regardless of what the specific shape of the deployment that you pick, whether it is thousands of agents collaborating very on a on a single problem, or whether it is multiple large agents that are all, you know, in very autonomous ways trying to talk to each other and, you know, arriving at a common business solution.

Whatever the case may be, you are armed with the best possible tools that can enable these agents to operate in an open, interoperable fashion and arrive at their, you know, arrive at the outcomes that they're trying to get to quickly and efficiently, and in a way that is secure and scalable. So that's the goal that we are driving towards. The name of the game for us is optionality in the space of a very diverse, heterogeneous, uh, set of, um, you know, types of agents available to you.

absolutely, and the there's a sense of many years ago as well, not not that many, uh, just a handful, uh, even after the chat GPT and generative AI explosion that we saw at the end of 2022. There was this sense of you either build or or you buy, right? And it was a little bit of a dichotomy, a binary between the two. And I know that phrase is still out there. The audience has heard me say this a hundred times.

They still call it build versus buy. But at this point, everybody's doing a little bit of both. It might be your build and buy ratio. And if you're familiar with the movie Mean Girls, I'm going to make this fetch, I'm going to make this a thing. But most folks who come on the show kind of recognize that reality that it's a little bit more of a ratio. Interested, especially in what you just uh what you last said, just about finding that balance between large and small models, large and small agents, really getting the the combination of tools correct.

How does that inform maybe the advice that you're giving partners just about building versus buying, building to buy, um, getting that ratio correct?

That's a great question again, and I feel like this is, you know, it's not going to be an either-or, right? Like people are going to build what they have to, they're going to buy what they can, right? Everyone wants to build a thing that differentiates them, that gives them a a leg up in their business, that provides value to their customers or enhances the the value of their product. And if there's something that's in the market, they definitely don't want to be reinventing the wheel.

However, when you start to do the thing where you have this combination of agents that are homegrown and agents that you're using that is running on somebody else's platform, that perhaps you have no idea on what framework it was built. All you know is the agentic interface. You don't even know, like, how exactly do I find these agents, talk to them, et cetera? That's where you need enterprise-grade tooling that allows you to do that with a degree of confidence and and with the security and observability and all of that built in so that you can do this with confidence, right?

And uh, we ourselves, for example, one of the multi-agentic systems that, you know, perhaps I I just, I mean, uh, alluded to in passing. One multi-agentic system that we built out here at Outshift is a uh, is a is a healthcare application that we built, which is it has, you know, multiple different agents. There's one agent that's running on our own WebEx platform. It interacts it interacts with the end user, understands what their symptoms are, does an initial triage, and then it talks to agents that are running on completely different platforms, that are running on healthcare platforms and insurance platforms.

So it'll do things like, well, it'll talk to the insurance agent, and it will understand that for this specific individual, what is the insurance coverage that they have? What are the types of services that are available to them? Who's in network, who's out of network, you know, et cetera, et cetera, what's their budget, and so on. And then, based on that, it then understands the optimal healthcare provider to reach out to. And then it uses another agent running on a completely different platform and goes out and makes an appointment for them.

This is an a a multi-agentic system that we built, and that's actually we've demonstrated it and so on. But it's a great example of an application where agents are running on completely different platforms. They, uh, and you know, they've discovered each other through a common standard interface that we've we've built out. And then they've been able to communicate and address a a problem. And we've now built a multi-agentic system where the whole is definitely much greater than the sum of the parts, right?

And this is the kind of thing that we want to enable for all enterprises. So this question of build versus buy, it is a spectrum, it is a it is a continuum. And regardless of where you fall on that spectrum, where the dial is for you on that continuum, we are trying to give you all of the tools that you need to make it possible for you to pick the best possible agents that are there. Build what you need to, buy what you can, and then stitch them all together at enterprise grade.

Absolutely, and I, and I still think leaders are trying to find that difference, especially now that it's more of a ratio necessarily than a binary. What should leaders look for instead of buying more tools? How do they know when to hold back, especially with all the marketing around?

So this is one where I don't have a formulaic answer because again, like I said, this will be very specific to the needs of your business, right? What I would, there are some criteria that I would use, some guidelines that I would definitely use. This is a time for people to be experiment, you know, this is a time. If you're an enterprise that has not yet built out a real complex multi-agentic system yet, I would say, what are you waiting for?

You need to dive in. This is the time to experiment, you know, the the field is moving fast, and if you do not experiment and try out things, you risk get getting left behind. So you need to experiment. But at the same time, you need to recognize the fact that the state of the art is changing rapidly, right? Like in this field, what was, you know, what, what happened last week is already old news this week, you know, especially, I mean, it seems to be getting ratcheted up even more and more.

And we saw all the stuff that's happening with, you know, the cloud bot became, you know, Mort bot became open claw and the moat book and all of that. You know, it's just like there's a new thing coming every day, right? So, uh, that on the consumer space, of course, as we talked about on the enterprise uh side, things move with a little more deliberation, thank God. But regardless, it is moving very fast and you need to move with the times.

But you need to do so in a way that doesn't compromise your business, that doesn't expose your customer data. So you need to recognize that this is we have to experiment, but at the same time, we have to do it safely. And we we may we have to recognize that some of the decisions we make now, we may have to revisit later. So, you know, something that we build now, you know, we may have the ability to, like, use a a canned service later that does it better.

But regardless, we need to have the the tooling in place that allows us to do that, right? Like, so, again, for me, the name of the game is optionality and being able to try different things without getting boxed into a corner, thinking that this is the one way that it will happen because that is likely to change. So I think, as an enterprise IT leader, that is something you need to think about. So for us, it's always been about openness and interoperability.

So whatever tooling we build, we are building it out in the open, we are building it in an interoperable fashion. As new standards Emerj, we've been we've been adopting them or collaborating with them, or interoperating with them. And the goal is to enable this open interoperable, what we call the Internet of Agents, now we call the Internet of Cognition, being able to allow agents to talk to each other, communicate, collaborate in that fashion that doesn't prickle things, you know, new services from coming on board and still being part of that overall ecosystem.

So that's kind of the the name of the game for us. And I would encourage IT leaders to also look at it from that sense.

Absolutely, we we could do a whole nother episode, uh, about IT leaders. And you know, hey, uh, have your people talk to my people, we'll have you back on the show for it, as they say. But you you're bringing up a a number of points that that we've heard on the show before, or to sort of nutshell these into a a couple of different phrases that that I've for choruses that I've heard for from guests on the program.

Jump, uh, don't wait to jump in the pool, but get in the shallow end. Don't jump in the deep end. And yes, you want to start with an area that's that's experimental, but still important to the business and able to be sandboxed, able to be a safe place for experimentation and failure. But you still want to be doing something that's important to the business. I believe, uh, the ratio we've heard on the show again is you want to do 10% of an important problem very, very well, which gives you 90% space to mess around, get your hands dirty, be that kid with the play-doh and and and really, really understand these systems from a tactile perspective.

This last question, maybe for folks that they're past that phase, they're in the experimentation phase, maybe they're finding themselves on the wrong side of that ninety-five number that we heard from, uh, MIT. That's that's just making shockwaves through every conference I go to at this point about AI. But let's say you're in that experimentation phase and you're blocked on the progress. What is your perspective and advice, uh, for those folks for not, you know, they've jumped into the pool, they've they're not letting this moment go to waste, but they're finding that this, at least this process, might not have been as simple as they first suspected, even if they went in with a healthy amount of skepticism, as we all

recommend on the show.

Yeah, w- what I would say is stay the course, you know, yes, these things take longer than than you might have initially envisioned. Um, and sometimes when you see the uh, pace at which you you hear these new announcements happening in the consumer space, it's easy to get disarmed to say, why aren't we moving faster? I I would, uh, encourage them that, you know, enterprise, they are doing it the right way. If you think that you're moving more deliberately at a at a in a more intentional way, that is great for your enterprise.

Like Matthew said earlier, you might want to take a look at have you picked the the lowest hanging fruit here, right? We did have, I mean, this is a playbook that we've seen in the past. If you thought about if you've been through the cloud migration um journey before, then that first when cloud became the deployment option for enterprises, it wasn't that they jumped in and took their, you know, business-critical applications and moved them to the cloud, right?

That was not how it happened. People started picking up low-hanging fruit, they started picking up easier, more tractable uh, workloads that can be moved to the cloud and gain confidence in that, gain the operational expertise before they started lifting and shifting. And some of them never did, right? Like, again, to my earlier point of AI being the same way and that there will be a heterogeneous, you know, diverse set of deployments. Some workloads, some AI agents are never going to be, you know, something that you outsource to somebody else.

You would want to build it yourself and run it yourself and observe it yourself, so secure it yourself, et cetera. But we have seen this before. There is a playbook here. So you want to pick off the lowest risk workloads that you have, the lowest risk problems in the enterprise that are most amenable to being attacked with AI, that that an agentic approach will solve. So spend some time thinking through that, pick off that low-hanging fruit, build up your operational, you know, muscle on how to deal with these complex systems, how to manage them, secure them, how to observe them, et cetera.

And then once you've gotten a a a degree of expertise and and competence at managing these systems, then you can start, uh, attacking the harder problems. But this is not the time to be taking your foot off the gas. You you want to continue experimenting and iterating as fast as you can because this is a big transformational wave and you do not you want to be riding the wave, you do not want the wave to be crushing.

And this answer can very much depend on organization to organization, industry to industry. I still think it's very important, especially when we hear an answer like that where you want to be looking for the lowest risk task, you want to be looking for the lowest hanging fruit. I do worry the audience might go home thinking, Oh, I need to pick something that's not that important. And I think you actually need the the task that's the most important with the least amount of risk with with the that's the easiest to do.

And any advice, maybe for trying to find that that sweet spot between easy, low risk, but also important to the core of the business.

If I had a dollar for every time I've been asked that question, everybody is looking for that answer, you know? And I would say, you know, think about what are the things that, what, you know, your core business is, what are the different activities that your business is involved in, and then think about what are, like, the for the agentic capabilities that you have in terms of information synthesis, in terms of being able to make sense of vast amounts of data, in terms of semantic analysis and things of that nature.

Where are those Venn diagrams intersecting? Where are the things that AI and agents are best at? Where can that make the most impact for my organization? And then, of course, use the lens that Matthew just said earlier of like, of course, it has to be lower risk. You don't want to be exposing your your customer data or PII data or what have you, you know, compliance or regulatory, uh, data that is that falls under those regimes.

Do that, but taking some low-risk but high-impact activity, considering the necessities of your business and the capabilities of AI and finding where that intersection is. I won't be able to hazard a guess precisely on what that would be for your specific business, but there are principles here that you can bring to bear on that and kind of figure that out. And then pick off a few and try them out and see what, what kind of, and then measure, right?

Like, it's always about iterating, measuring, improving, and and then rinse and repeat. So basically, measure what the out what your outcomes have been, uh, with the use of AI, and then branch out from there. So that's what I would say. I know it wasn't a very specific answer, but that's that's the best that I can get.

It it's a little bit it it's a little bit of a Rorschach test. And and if we have you back on the show and we do this in person, I owe you a dollar. But it it really is, I I think, a worthwhile question to ask, even for the the diversity of of answers in that. Yeah, there might not be a a specific one, but I think, especially for this area, where even the executives that have been at this for for a decade, even more, I I don't think they're officially settled on whatever philosophy they have, uh, for that selection.

They just have their best guess and every answer is different. Yes, I probably owe, uh, quite a number of executives some dollars, given how many times I've I've asked, but I think the audience always benefits, and I really appreciate you obliging me today. Papi, thank you so much for being on the program. It's been a great pleasure having you.

Thank you so much for the great question, Matthew. It's been a pleasure.

Wrapping up today's episode, I think there are three key takeaways from our conversation with Papi. First, enterprises aren't held back by model performance, but by the lack of a shared cognitive layer that lets agents understand context together rather than operate in silos. Second, scaling requires embracing heterogeneous agent environments and prioritizing open, interoperable foundations that prevent lock-in as the technology evolves. Finally, real progress comes from choosing low-risk, high-impact starting points that build the operational muscle needed for more complex multi-agent systems.

A quick note for our executive listeners. Emerj invites enterprise leaders who are driving meaningful AI initiatives to share what they are learning with a peer audience. If you're moving real projects forward and want to be part of that conversation, you can learn more at go dot merge slash expert. That's G O dot E M E R J dot com slash E X P E R T. For further executive-level analysis and to join our network of leaders delivering workflow impact with AI, visit Emerj dot com.

On behalf of the team at Emerj, we'll see you on the next episode.