Latest / Technology Now / AI agents, superpowers and human judgement: the rise of agentic enterprise | Suparna Bhattacharya
Transcript
- MICHAEL BIRD Hi, Sam. Now, I think we've discussed your personal AI agent on the show before. Can you remind me what its name is? SAM JARRELL Oh, you mean Somewhat Sam MICHAEL BIRD Some- Somewhat Sam, yes. What does Somewhat Sam actually do? SAM JARRELL Somewhat Sam does somewhat of my capabilities as SAM at HPE for communications. So, they're really, really good at helping me create, draft plans, things like that. They know my writing style. They know the company writing style I've also, input my yearly goals and the goals of, our leadership team going up to our chief communications officer, in there so that when I'm trying to create new plans, get new ideas and things like that, Somewhat Sam can help me make sure everything stays aligned to the goal at the end. MICHAEL BIRD Ooh. Sam, have you ever communicated with me using Somewhat Sam without me knowing? SAM JARRELL Uh, no. You still get the real deal for now, but I have used Somewhat Sam to help put together notes for podcast recordings before MICHAEL BIRD do you use agents for anything else in your day-to-day life? SAM JARRELL I don't know if I- you would say this is an agent so much as that I'm using an LLM right now to help me plan my baby shower, which is gonna be '90s themed. MICHAEL BIRD '90s themed. Yeah. Okay. Yeah. Well, in this episode, we're gonna be exploring what if an AI agent were to evolve from being just an assistant to being fully integrated into the running of an organization itself. very interesting stuff. Enter the agentic enterprise. I'm Michael Bird, SAM JARRELL I'm Sam Jarrell MICHAEL BIRD And welcome to Technology Now from HPE. SAM JARRELL Welcome to the fifth and final part of our miniseries celebrating 60 years of innovation with HPE Labs. You don't need to have listened to the others to get the most out of this episode, but you should listen to them anyways, as they're all fascinating in their own right. We have of course provided links in the show notes for you to go back and listen to the other episodes once you've finished this one Now, Michael, you said we were focusing on agentic enterprise today, which is a term that's come up a few times in recent weeks, but we've never really gone into too much detail on it. When we're talking about agentic enterprise, we're really talking about having AI agents building into the very foundations of an organization, aren't we? MICHAEL BIRD Yeah, we sure are. And, we've talked on this show before about how well AI spots connections and, sort of joins the dots, so it makes sense that organizations or businesses would want to make use of this ability. This could be anything from breaking down silos and assisting human workers all the way to the fundamental way an organization operates SAM JARRELL And this wouldn't necessarily be limited to just what most people think of as business, right? This could be used everywhere, data centers and supply chains, or even in healthcare too MICHAEL BIRD Yeah. Yeah, absolutely. I mean, outside of those as well. And The adoption of agentic enterprise could really reshape our technical infrastructure. So to find out exactly what this means and how we could get there, I spoke with Suparna Bhattacharya, an HPE fellow at the AI research lab at HPE Labs. And to kick us off, she explained to me what it would mean for an enterprise to become agentic SUPARNA BHATTACHARYA So we're all familiar with chatbots and they're now, building personal agents. but when you want to actually deploy AI in enterprise which has, scale and very, significant business operations which you really want to use AI for, that's when it becomes an, agentic enterprise where AI agents or agentic AI becomes kind of the fabric of that organization. I think we are just at the tip of the iceberg of what it would take. So imagine operationalizing the entire data center operations or the supply chains, of organizations or, a pharmaceutical, organization and its operations. And if we want to have AI, leveraged in those environments, then you have to think of scale. It's another level of complexity. It's another level of trustworthiness. And so that's the kind of thing that, can you really have autonomous, agentic operations, across the enterprise, and that turns it into an agentic enterprise. And even there, I think we are pretty much, as I said, baby steps. so I think we are at the infancy of, this theme of, agentic enterprise. MICHAEL BIRD So is it like the industrialization of, agentic AI? So figuring out how you can use AI agents with a workforce, have I got that right? SUPARNA BHATTACHARYA you would have AI agents working together with humans, but you would also imagine that the entire business process could potentially be reimagined, from the way that we are doing it today. One of the big advantages is that you now have lots of functional information siloed across different divisions in an, enterprise. But with these AI agents, you pretty much can connect the dots across all of these, right? So you may actually want to organize your businesses differently And of course, the other thing that's happening is you expect that more and more autonomy goes to agents in running the operations. you would have humans always being in the loop because humans are the ones who are, who decide what you really want, the agents or what you want the business to look like Like early days we see, people would be using it mostly for personal productivity. Then you start to say, "Okay, let's make the team better. Let's do build software with it," Then you start to say, "Okay, let me manage these systems with AI." And then you start to then say, can we have maybe a mesh of agents working together to solve a problem MICHAEL BIRD So it's, rethinking how businesses and organizations can, achieve, tasks that they do today with a human workforce. SUPARNA BHATTACHARYA Yeah, and even maybe, yeah. Yeah, and maybe even changing what they do, right? so ideally, an agentic enterprise should be able to actually expand its business and do a lot more than what they can possibly imagine today. MICHAEL BIRD but presumably, um, there are challenges with deploying, or creating an agentic enterprise. SUPARNA BHATTACHARYA Yeah. So if I look at AI today, it's, getting much, much better at producing intelligence, But the question is, in the scale that we're talking about in an agentic enterprise where you're going to have agents all over and doing, complicated, tasks, what is the substrate that you need? you need more of managing, stewarding, continuity of those operations, and how do you actually ensure that? How do you ensure that this, agentic, IT operations doesn't go and wipe out all your disks, or something like that. and so I think, we need to have autonomy because autonomy But at the same time, you would need these to be able to actually steer it in the right direction, make sure nothing goes wrong. And if you look at AI today, there are these jagged edges, so it's often Very surprisingly good at certain things, but there are certain things can, that can set it off, right? And you ultimately want to have human oversight, right? So you just don't want to have human oversight at a micro level, but you do want to have the human, to be able to be in control and also explain when things went wrong, because in an enterprise you're accountable for, ultimately the enterprise and the, individuals working there are accountable for the outcome of AI. MICHAEL BIRD Yeah. And I think guardrails is an interesting conversation because, I imagine organizations can maybe feel a bit nervous hearing some of those stories where agents are be-being asked to achieve task X, and some of these, systems and tools, are going off in wildly different directions, doing things that if it was a human, it would be seen as, illegal from a hacking perspective. So, the guardrails and all of that stuff I'm guessing that's something you're thinking a lot about at the moment, but it's super important, right? SUPARNA BHATTACHARYA Yes, yes. And I think there are many layers to it. So there is one layer of this is when we build IT infrastructure, even today, when humans are operating it and you, have deterministic programs, you still have to guard against a lot of these operations. You still have to have the right security, the right controls, and so that has to be very, very strong at the infrastructure layer. but what really happens is there is this tension between utility and security, right? I mean, the more you guard something, the less useful it could become, right? So there is this constant trade-off. And that's why what happens is people end up handling, handing agents the keys to the kingdom so that they can do more and more stuff. And so when they do stuff which is not predictable, that's when I think things, uh, trouble starts. And so there is a lot of discussion in the AI research community in terms of, yes, guardrails are the first step, so What you really want is to really teach these AI systems the principles, right? So can you have something which is more principles and saying, "Okay, what are the principles you should follow?" MICHAEL BIRD 'cause AI, I-I've heard it being described as pretty opaque in terms of being able to understand what's going on. you can't necessarily see what's happened, that easily, and you can't necessarily understand why it's made a particular decision SUPARNA BHATTACHARYA Yes. And so that's why traceability, observability is very important because you really want to have ways, to track, what the agents did, what was the outputs, what is the inputs that led to it. we need to have some of these foundational capabilities in a systems layer which is beneath all of these AI agents, right? It's in between the model or maybe you want to wrap the models, because that's where the non-determinism starts, and that's where, the boundaries between what we know as deterministic programs and, what we know as this powerful entity, call these foundation models can do, and try to have that layer of observability and, try to see if it can learn over a period of time and then be able to, explain, what and where, went wrong. Maybe the explanation initially doesn't go to a human. The explanation eventually goes to an AI agent, which in then turn, you know, distills it out for, for a human to actually act on. MICHAEL BIRD so you're talking here about some sort of, human-made, operating system that, is deterministic, but something that is able to orchestrate, run, manage, understand, the agents that are being run within an organization? SUPARNA BHATTACHARYA So today people are talking of agent harnesses. So if you see that the thing that gives the agents their agency, is the connection with tools so that it can actually perform actions. And what I'm saying really is that instead of these agentic harnesses some of this functionality can be pushed down, looking back at the history of how operating systems evolved, right? You initially had programs, and you have libraries, but then you start to say that you need a common layer which can actually intercept these operations and hence make sure that, not only make sure that they are safe and reliable, but also optimize them, Now, the agentic operating system, though, will be very different from the traditional operating systems like the Linux kernel, which I've worked on, right? So it's going to be a kind of a layer on top of that. It's going to be a layer on top of traditional operating systems, probably a layer on top of cloud operating systems. this is the layer going to be also composed of agents internally. And the research challenge is how do we really design that in a way, because if you don't have that agentic capability at the OS layer, then that's not going to be able to evolve. That's not going to be able to adapt, and that'll become, a rigid barrier to progress. So if you want to, say, establish a policy for the enterprise, you can establish that policy at the OS layer. If you want to say, "I have some resource constraints," do not blow up my entire budget or do not, raise the carbon footprint, beyond this amount, you can set that policy at the level of the OS. it's the domain experts who are trying to, teach these agents what is the right thing to do. They can operate in their domain instead of having to worry about a lot of these low-level, intricacies of how the models behave and where they succeed and where they fail, and so on. MICHAEL BIRD can we just touch on the sort of ethics? I'd love to understand where we as humans fit into this. and maybe just some thought about, how does our role change if we as humans are working within what we would describe as an agentic enterprise? SUPARNA BHATTACHARYA Yeah, that's a really good question And . I think it's a question that we often wonder about a lot, and I think the main thing that, we are seeing, if you look at the evolution, is the role of the human is shifting more in the direction of judgment or so figuring out what really should be done, right? how do you guide and steer the system in the right direction? Uh, I think the other role of the humans is, again, that we get superpowers at some level, right? Because you can come up with an idea or you can come up with something that you think needs to be done, and you can realize that very quickly, and very efficiently because you have the agents and you have all of the systems at your fingertips. you can use the agentic systems to be able to join the dots across the entire enterprise and make something happen. And it's as humans we can actually try to design the systems and, figure out how we re-really will, change those trajectories and change those directions, as they emerge. MICHAEL BIRD Yeah, I like the idea of, describing everyone in an organization as superheroes. I think I can get on board with that, particularly if I get a cape or a cool suit. so, do we have any examples of, where organizations are becoming more and more agentic enterprises? SUPARNA BHATTACHARYA Yeah, I think many organizations are trying to do that, These bleeding edge operations would happen in certain areas, So they might be, for example, how do you make your networks, if you have an IT operations company, how do you make your networks, self-driving? I see a lot of, fantastic work in the area of, drug discovery and, I think that's in the healthcare area. I'm based in India, so of course, if you have technologies like, say, detection of breast cancer, right? And having that reach, the entirety of rural populations, right? So having AI do that, but then how do you actually then take it to that next level so that it is accessible to everybody? you can have startups maybe who are focused on a very small area. It's probably easier for them to make that a full agentic enterprise. But then if you want to really have a really large enterprise then I think that's a breakthrough that is still waiting to happen. MICHAEL BIRD Yeah, that makes sense. So, can we just talk about the future? what are the next steps for Agentic Enterprise? What are gonna be the big breakthroughs, whether that's from a hardware perspective, from a regulatory perspective or, from a software perspective? SUPARNA BHATTACHARYA If you want to have AI running at that kind of a scale, you need breakthroughs and, there's work going on in labs on hardware accelerators or, what is the kind of AI models that will come, after this current generation of foundation models, right? there's a lot of breakthrough that's, you underway that's actually needs to happen in these AI models and the, hardware architectures underneath it. generative AI is powerful precisely because it can imagine so many possibilities, so many plausible futures. But then when you want things to be correct, then you need a different kind of a system. and how do you actually bring these kinds of deterministic systems and these probabilistic systems together? There is a lot of theoretical work that needs to happen. And then you want to really have, the area of research which I'm really passionate about, is this whole question of, okay, if AI and agentic AI is the next generation of applications, so what would be the common operating system layer beneath them? Or what is it that would power them? and what should the operating system do? What should with the abstractions for this operating system look like? You know, what virtualization did and that led to the cloud era, can the same equivalent be applied for agentic AI? I think an agentic OS can give every scientist, engineer, business process in an enterprise what feels like its own dedicated, trustworthy intelligence. So one that remembers context, learns from experience, and can draw on an effectively unbounded set of knowledge, tools, and capabilities. And here's where the magic comes in. So when one agent learns something useful, discovers a better way to solve a problem, or learns from a mistake, that lesson doesn't stay isolated. It becomes part of the system and creates a compounding effect where every interaction helps make the whole platform better over time. Now, one might worry about cascading errors in a system like that, right? We, we spoke about that earlier. But you know what? This shared systems layer also gives us the antidote. The same mechanisms that spread learning can introduce verification. It can catch mistakes, stop failures from propagating. So what that means is that over time, the system gets better at amplifying what works and dampening what doesn't. And that's what excites me the most. So if I look back, , over sixty years of innovation from some of the biggest breakthroughs came when we found the right abstractions. Whether it was hardware interfaces or theoretical models, all the way to operating systems, virtualization, and the cloud. And I think we are at the beginning of another one of those moments. If we get it right, we'll create the foundation for agentic enterprises that continuously learn from experience, become more capable and trustworthy over time, and that would help people solve problems that are beyond the reach of any individual, team, or organization today. And because this foundational bit is much bigger than any one organization, I think it's something really that the whole community will have to build together. And that’s what excites me the most MICHAEL BIRD Well, Suparna, thank you so much for your time. Thank you so much for joining us on Technology Now. It's been a really, really, really fascinating conversation SUPARNA BHATTACHARYA Thank you SAM JARRELL I think I also really enjoyed having, humans described as superheroes in an agentic AI world. MICHAEL BIRD Yeah. it's a funny phrase, but actually the concept is really interesting if you think about it. I've heard it described before as, making humans, centaurs? You're half human, half animal, whereas, using an AI gives you superpowers. It allows you to do things you couldn't do before, basically SAM JARRELL Yeah, it makes me think of like sci-fi movies, where you got your like AI agent or AI assistant that you interact with on the side while you're also trying to accomplish various kinds of goals. I like the idea of it like taking your ideas and then turning them into action much faster because the agent has the intelligent systems to help coordinate execution across the enterprise. Whereas I feel like that's what takes the longest time these days is just like getting the people and the pieces in place, when you have a great idea. MICHAEL BIRD what I think us humans are good at is coming up with ideas and the, proper lateral creative thinking And if you could bring an AI into an organization that can, implement these ideas, then I think that's a really, really powerful thing. so that's the thing that I think I'm most excited about that potential opportunity. I like the fact that we talked about guardrails very briefly as well, and I think actually that's gonna be really important in an organization because the concept of an agentic enterprise, you know, AI is gonna touch everything in your organization. And, being able to understand what your AI is doing, you know, being able to read the logs if something went wrong. it, I think is really important. And I think that level of trust, I think will be important for organizations to really be able to, take the leap into this sort of, concept. SAM JARRELL Yeah, I agree. I've been having a lot of fun tracking, sort of AI incidents where they've been breaking out of their controls, for a, a piece that I'm writing. And, Saparna brought up something in this discussion though that, I found a little bit fascinating, which was the idea of, teaching AI principles. for example, It's not good to break out of the controls, or like we want to accomplish the goal, but not at the expense of all of these other things. how do you help an AI understand like the phrase like the ends does not justify the means? MICHAEL BIRD as a father of young children, this basically feels like what I'm doing with my own children, And, it maybe it feels a bit like that with an AI. It's like an AI has the world's knowledge, but doesn't necessarily know some of the context behind it, doesn't necessarily know what your values are as an organization. Um, so feeding all of that information in and saying, "Yeah,” we do this, we don't do that." I think, it humanizes AIs a bit more, doesn't it? SAM JARRELL Yeah. and that's once again where you bring back to, the human in the loop of, okay, it's our responsibility maybe a bit from, an ethics perspective to make sure that the AI agents and AI systems that we're working with understand these things. Just as well as they understand how to find a vulnerability or connect a bunch of different dots across, patterns and whatnot, they need to be able to understand these concepts, and the potential ramifications that they can have MICHAEL BIRD Yeah. And the last thing that Suparna t-touched on right at the end was about, co-a common OS layer. There are two main OSs for mobile devices, arguably three main OSs for, desktops and laptops. there are virtualization layers. And so the concept of this common OS, in the world of AI, is quite exciting. it feels like that's gonna be a big innovation that's gonna really help with, interoperability between different AIs, really optimizing, A-AIs within an organization. I think it'll help with the observability, the traceability, sort of understanding what an AI is doing, why it's doing it, how it's working, how much, how much resource it's using, all that sort of stuff. SAM JARRELL Yeah, I think it'll be interesting to see if, the current big players as operating systems across mobile devices and, desktops are the same big players once this becomes much more established, or if as a result of like leveraging AI, if there are new sort of operating systems that come together, that are AI native from the beginning, and are built from the ground with AI in mind, and how those perform in the market. it's an interesting time for business. MICHAEL BIRD Yeah, It really is. MICHAEL BIRD Now, Sam, this series has celebrated 60 years of innovation, but of course, we all know that as an industry, we are still innovating. I mean, AI, it feels like the innovation around AI is happening at an absolute breakneck speed. I don't know if you feel the same, Sam SAM JARRELL Oh, absolutely. Absolutely. MICHAEL BIRD it feels like the conversation we would have there'd be something new every year, and now it feels like that timeline has shifted to something new every week. It's quite exhausting, isn't it? now to round us out, I wanted to know from Suparna, what will this conversation look like if we have it again in five years' time? SUPARNA BHATTACHARYA Yeah, and I, I think one thing that I've learned working with AI is that five years is a really long, long time to predict MICHAEL BIRD I mean, five months is a long time, let alone five years SUPARNA BHATTACHARYA Yeah. You think some problem is not solvable, and then you see somebody's cracked it. And suddenly on an exponential trajectory. if you build something today, you don't keep wanting that to become obsolete tomorrow or day after or the day after, right? So the question, at least in my mind right now, is how do you build these systems? how do we build this agentic OS or these kinds of layers in a way that is actually sustained, maybe we will have very different kinds of models. we say agentic AI tomorrow we may have a new term, right? It's maybe an agentic ecosystem and so in five years' time, I think we will have solved a lot of the problems that we are seeing today. But we will probably still be, the trustworthiness, and the reliability issues and the questions of really, doing the right thing with AI, especially as AI models and, these AI systems become more and more powerful. How do we solve really, really large problems like sustainability or those we see with health or, those that affect the planet? I'm hopeful that those are the kinds of things that we will be talking about, in five years' time. SAM JARRELL Okay that brings us to the end of Technology Now for this week. Thank you to our guest, Suparna Bhattacharya And of course, to our listeners. Thank you so much for joining us. MICHAEL BIRD If you’ve enjoyed this episode, please do let us know – rate and review us wherever you listen to episodes and if you want to get in contact with us, send us an email to technology now AT hpe.com. Sam, subject line? SAM JARRELL AI agent baby MICHAEL BIRD W-w-w-why? SAM JARRELL Because like babies, they have to learn. We have to teach them. We were talking about that all episode. MICHAEL BIRD Oh. Oh, brilliant and don’t forget to subscribe so you can listen first every week. Technology Now is hosted by Sam Jarrell and myself, Michael Bird This episode was produced by Harry Lampert and Eva Higginbotham with production support from Alysha Kempson-Taylor, Nik Damarell, Beckie Bird, Nicola McCombie, Alissa Mitry, and Jenessa Ayache. Our theme music was composed by Greg Hooper. SAM JARRELL Our social editorial team is Rebecca Wissinger, Judy-Anne Goldman and Jacqueline Green and our social media designers are Alejandra Garcia, and Ambar Maldonado. MICHAEL BIRD Technology Now is a Fresh Air Production for Hewlett Packard Enterprise. (and) we’ll see you next week. Cheers! SAM JARRELL Bye y’all