09x09: AI Gets Personal with Agents Acting on Our Behalf

Speaker A: Agentic AI is an autonomous system that learns, adapts and uses tools on behalf of its users. This final episode of Season 9 of Utilizing Tech brings hosts Stephen Foskett, Frederick Vanhaeren and Guy Currier together to reflect on the lessons we've learned over the last few months. Welcome to Utilizing Tech, the podcast about emerging technology from Tech Field Day, part of the Futurum Group. This season focused on the practical applications of agentic AI and other related innovations in artificial intellig. I'm your host, Stephen Foskett, president organizer of the Tech Field Day event series. And joining me for this final episode of season nine are, uh, my two co hosts from the season, Frederic Van Haren and Guy Currier. Frederic, Guy, welcome to the show.

Speaker B: Well, thanks for having me. I'm Frederic Van Haren, the founder and CTO of Hifens and we provide HPC and AI consulting services.

Speaker C: Yeah, it's great to be here. Guy Currier. I'm an analyst at the Futurum Group and um, I occasional participation and participant, uh, in Tech Field Day as well.

Speaker A: Absolutely. And I'm Stephen Foskett, organizer of Tech Field Day, including the AI Field Day event that the three of us all attended here during the recording of this season of the podcast. And, um, going forward, host of the new uh, utilizing AI podcast over on, um, TechStrong AI. But of course we'll be back with future seasons of Utilizing Tech as well. Uh, let's sort of, I guess wrap up season nine here and talk a little bit about the lessons that we've learned rather than making this just a retrospective of the various guests that we've had this season. Uh, let's talk about some of the takeaways. Um, Frederick, I'll start with you.

Speaker B: Yeah, I think we learned a lot. I, um, think agentic AI is still kind of a moving target in the sense when we ask people about a definition, the definitions can vary a little bit. But I think overall people have a great understanding or a better understanding, um, of what AI can do as far as, um, reasoning and thinking, um, with large language models. And I think that's what we saw during the episodes. I mean we had some people talking about applications, we had some people talking about use cases. Um, I think overall, uh, in combination with AI Field ah, Day, I think we got a, at least I got ah, kind of a view on what people are doing on both sides of the fence, customers as well as vendors.

Speaker C: I think, uh, one of the things that has most impressed me, um, is, um, how rapidly it's developing. Everything in AI is developing rapidly. Um, in fact, it's probably changed significantly since we started this podcast, um, series. Um, so that's one, um, I think from. As a practical matter though, there's so much that you can do just to step into, uh, to move sort of, let's say beyond, certainly beyond chat, certainly beyond basic use of AI and into Agentic. It's a question of connecting Agentic to systems and to each other and doing a little design work. It's not necessarily that far. You don't have to do the latest and the greatest. And there are a lot of interesting platforms and ways to, to do this. There are a few AI studios out there to help you build agents, um, some that are incorporated with maybe services that you already have.

Speaker B: Yeah, Guy, I think that's a, that's a good point. I also think that applications and use cases typically came from vendors, or at least the view from vendors. I think Agentic, AI and mcp. And like you said, the studios that people are delivering to the market are helping people that typically were not engaged or at least not on the vendor side, now can build applications that are kind of very close to solving their problem as opposed to applications that are solving other people's problems.

Speaker A: Yeah, that's a good point. And that came across of course on our episode with Articulate, but also throughout the season when it comes to agentic. If the idea is that you're going to make. And again, we spent a lot of time trying to define Agentic earlier in the season, but, uh, let's sort of roll with that. If the idea is that you're trying to make AI agents that can act autonomously, that can ingest and process data, that can call other tools on your behalf, it really is important to make sure that they are up to the task, that they're not just sort of, um, generic, and that they are able to respond to the needs not just of the business as a whole, but of the users, the people who are trying to make use of those agents. And that came up, um, you know, many times throughout this. I see a really strong parallel between um, the process automation space and the agentic AI space in that in both cases it's sort of, sort of a, um, twist on that whole, no code, low code concept or, you know, a way for people to make AI do things on their behalf. And it's funny because during this recording of this season, um, a couple of things happened. One, um, as I talked about at AI Field Day, the uh, Futurum m Group released a report on um, enterprise agentic AI platforms where they highlighted some big companies, you know, Salesforce, Microsoft ServiceNow, IBM, those kind of companies. Uh, at the same time uh, we also saw people really leaning into things like the Perplexity web browser and um, the new OpenAI web browser as a way to basically have a personal Agentix system. And at the same time of course Apple is hopefully going to be releasing more AI features on iOS and Google just keeps pushing Android forward. And all of these I think reflect that, that idea that that Agentix should be personal, should be usable, should be um, something that people can really interact with or else it's really not going to be able to achieve the goals. Right.

Speaker C: One of the things I definitely picked up from the series specifically was when, when um, I did the, the episode with um, with Kami Waza. Because one of the structures that uh, uh, Luke Norris there provided um, was around three different types of. Was. Was a view on three different types of agents. Because there are ones that are, and I don't remember the types but there was one. There's ones that are responsive, there are ones that um, are autonomous or operate on their own. And then there's a third type I'm not remembering right now that was really helpful because m. One of the mind shifts that I um, you know, have experienced um, since we started this uh, was as to what um, agentic AI is, what an AI agent is. Because um, I, I think just you know, kind of to be a kind of classic about it or something. I started off by just saying an AI agent is just an agent, it's just got AI in it. Um, maybe oversimplification but seemed to me that if you just think of it as an agent that uses AI then that helps, it helps you understand what it is that you would do with it. Um, that was clearly wrong. Um, because the way you build it, the way you run it, how it works is different than a standard agent. It's still an agent, it's a kind of agent, but it's different enough that it's not just an agent. That includes AI. Definitely not. Um, it can do things on its own, it can learn, um, which is to say it can be self trained. Um, it can interact in ways that are um, not just unpredictable but uh, what do they call it? Non determinative or something where it might, might do something different the second time. Yeah, uh, that is different from every other agent we've ever computed. Uh, or sorry programmed and used.

Speaker B: Yeah. What I think is nice about agentic AI is that we had access to large language models in the last couple of years. But the big challenge was how do you integrate all of these components with your own data? Right? Because in the end it's your own data that makes or creates the value of an application. And I think that's one of the things that Gentek AI and um, maybe, you know, maybe not necessarily a definition but it's, it kind of allows you to bring those different large language models together with your Data through standards, APIs if you wish. And I think on top of that that makes it very accessible to individuals. I think if you look at people building applications around large language models, in the past those applications were very static, meaning the large uh, language model couldn't change, you couldn't interact with another large language model, you couldn't daisy chain large language models. And today you do have these capabilities. And it brings a kind of an interesting factor to the foreground, which is an application today is more dynamic than ever. In other words it's not a um, finite state when you build an application it's a work in progress. And you can see that too with uh, with Vibe coding where they're not suggesting to build or create a prompt that defines your whole application, they're basically saying do it in individual steps. And that's really interesting because I believe that that allows you to build applications that, to Steven's point are more personal because you can iterate through it and you can start with the baseline and then add functionality. To tie this with, um, one of the episodes we had is that I, from a speech background, I always think about text as being the main communication piece. Um, there is multimodal, right? This audio, video, uh, text and this input and output.

Speaker C: What's really interesting, um, you know, uh, is this idea of. So I mean do we all remember when Google started uh, um, uh, publishing services, new services of various kinds going back 20 years or 15 years, whatever it was in beta and the beta lasted forever. And this was kind of may not have been Google leading the way, but it seemed that way to me. The perpetual betas, what you're talking about is true perpetual betas almost, it's almost like the agent, the software is never done. It's, it's always going to be. You can prod it into evolving, it can evolve on its own. It's never static anymore. That's ah, a kind of a wild concept. It's a little bit what Satya Nadella was referring to a year or so ago when he said there was not going to be SAS anymore, which is to say, you know, the creation of agents, creation of software, and then it's their destruction or they'll scoff into a hole until you need them again. Maybe you'll never need them again. Um, that's a very different way to interact with systems. Very different.

Speaker A: Yeah, exactly. And I definitely see that dynamism happening here in a way even beyond non deterministic. Um, it is almost as you're saying, both of you are saying, that the application you use or the workflow you use might be different to today than tomorrow, than the next day. And it's interesting, right before we launched this season, OpenAI introduced chat GPT5. And one of the hallmarks of GPT5, as I actually said back on the first episode, is that it is, well, I guess depending on how you want to define it, is almost agentic. And I've been doing a lot more work with GPT5 recently and um, it is really interesting if you watch the workflow there, how it interacts with you and to both of your points, essentially, you know, you ask it a question and it is calling uh, specialty tools to answer your question. It is responding to you by looking up data, doing a web search, using a calculator, you know, using Mathematica, using um, you know, variety of different ways to produce what it is that you're asking for. And specifically, you know, I've been having IT process, uh, to Frederick's point about multimodal data, I've been having IT process images and um, and give me JSON. And it is wild to watch that workflow because it is using, like I said, it's using Mathematica, it's using um, other uh, generative AI tools to process images and identify items in the images and it uses um, you know, all these different things. That's, I think, what we're looking for here with these next generation of tools. It's not about making an artificial super intelligence. It's about making a system that can really step through, you know, define the next phase, figure out the tool to use, use that tool, take that output, go to the next step, go to the next step. It's a very personal way to do this. Um, unfortunately it's also kind of frustrating and it's been kind of frustrating for me as I've been using these tools because I just want to shake it sometimes and say, no, you went in the wrong direction halfway through. But at the same time I feel like it's more likely to generate an answer than trying to come up with some super machine intelligence.

Speaker B: Yeah. And that's a perspective from a user side. Um, slightly off topic, but I watched an interview from the people that started uh, ah, Claude code and basically it was kind of interesting to listen to them. Everything is almost accidental. They had no intention of building it, but somehow they found something and then they build it. And another statement they made, which I found interesting, is that not only does an agentic AI system learn from its users, it's actually also learning from its own output. And that kind of tells me that we as consumers of agentic AI might not always understand what's going on. The people building agentic AI, large, uh, language models, even, they themselves have no idea where it's going. They're also being led by, by the large language model by itself.

Speaker C: I wonder if, though, just to spice things up a bit, if we're looking ahead. One of the issues right now in training is in training, foundational models. It's not quite an issue yet. It seems to be getting there, is that there is a peak, there's an optimal training level, um, in terms of quantity of data and number of cycles and stuff be because of the amount of available data. Um, in other words, we uh, could run out of data. As much as we talk about the explosion of data, we more or less run out of data in training the models and uh, the use of synthetic data or what have you. Or more specifically, we're looking ahead to where some of the data being used to train is actually output of AI that already used original data that was the output of humans. Right. So that's that decreasing quality possibility. This could get accelerated with agentic, um, what you're talking about, Frederick, is sort of systems of systems. And where the systems get abstracted far enough from sort of the original human origin, so to speak, they are still derivative. AI is still a simulation. Um, I tend to, you know, object a little bit towards like reasoning or thinking or that sort of thing or even the word intelligence because of it. And um, I'm just uh, I'm thinking ahead to how we utilize AI going forward in a way that remains productive, even if it starts to be a whole lot of AI talking to each other.

Speaker A: That's actually a really interesting point. I don't want to get all philosophical on y', all, but, um, we already did see some um, examples of AI agents talking to other agents and developing their own mechanisms of communication, their own vocabulary. Um, we are trying to use. I mean, I don't know about you guys, but I basically want AI to give me JSON if I'm using it in any kind of application, as an application agent assistant, that kind of thing. Um, I don't know that JSON is the optimal format for agents to talk to each other. I mean, with mcp, you know, again, we're trying to impose our human will on these things. I would not be at all surprised if future AI agents interact with each other in a, an API and exchange data in a format that is, I don't want to say completely illegible, but at least not what we would have designed. Because it turns out that that's an easier, better, more efficient or just sort of evolutionary sort of way of, of exchanging information. Um, because essentially if we're going to set this stuff out there doing things on our behalf, we've got to let it do its thing and you know, we can't micromanage it and babysit it.

Speaker B: Right. I think as far formats are concerned, I mean JSON is, is a very good format and you can pretty much communicate any type of data you want. The challenge with JSON is, is that JSON is not really meant for large amounts of data. I think the challenge becomes when you want to exchange a lot of data in a small amount of time, like for example, uh, cars driving and exchanging videos with each other. Um, that's where I believe Jason wouldn't do so great. But yes, uh, you know, JSON or something else, um, there is definitely room for uh, some kind of more advanced format, but we have been through so many iterations. I don't think JSON is a bad format at all.

Speaker A: Yeah, JSON is certainly the worst format apart from all the other ones and God, at least they're not using xml.

Speaker C: You're channeling uh, Winston Churchill there for us, aren't you? Um, yeah. Um, I think though, Frederick Stevens point as I took it, or maybe I'm taking it a step beyond is, uh, these systems are going to start designing their own interfaces to talk to each other.

Speaker B: Right? I think that's the next step, right? Is that the machines decide how to communicate with each other. I mean the bottom line is as long as the, the uh, communication channels are authentic, um, and follow certain guidance, maybe they will. Who knows, right? I mean JSON is still a human readable format, right? I mean the reality is though, machines don't need text, right? They need binary. And so they could talk in four bit as opposed to eight bit or whatever, right?

Speaker C: It's, or in who knows what it'll be like reading a uh, machine Ballot in Texas you won't know exactly if you're getting to vote for who you thought you voted for. I'm making a joke. It's a reliable system. But I do object to using barcodes to vote. There, I just said it. But uh, yeah, who knows how they want to talk to each other. They'll find some optimal way to them or in fact they might find a suboptimal way to that really doesn't work but that they came up with because they're non deterministic.

Speaker A: So let's kind uh, of turn the page and talk a little bit too about some of the other aspects. Um, you know we talked earlier about the platforms uh, that are being used. Um, part of the conversation that we had this time around was talking about um, running uh, agent agentic applications on personal devices um to the cloud. You know Frederick brought in the concept of multimodal data. Uh, let's talk about some of those other elements that are evolving AI. And again I feel like overall it's all about making it more useful, more personal, more um, actionable. So uh, what's your take on I guess the um, agentic platform concept? You know we, we've heard the enterprises um, embracing things like the Salesforce, uh, Agent Force platform. Um, you know we've heard companies talk about a variety of um, you know really kind of nuts and bolts, uh, almost um, you know, VM manager kind of platforms for running these things. We, we've talked about as well, um, the different ways that Apple and Google are evolving their ecosystems to run agents, uh, or AI, uh in instances locally as well as in the cloud. Um, where is that all going? What is the sort of the common through line that you're seeing in platforms to run AI applications? Guy?

Speaker C: Well, I think that you're giving me an opening to rant about data platforms. I think there are three platforms we're talking about and they can, they can integrate, you know, they can be presented to uh, to the user, um, just as a single platform. Um, but there's, there's the word that strikes me in terms of agentic AI activity is dynamic. More so than applications, more so certainly more so than, than model based inference even including rag, um, the type of data, the type of access required, uh, the types of, let's call them queries or needs. Data needs um, for an agent, um, will vary considerably. Uh and the data platforms need to be able to keep up with that. Not just from a quality standpoint, an availability standpoint and all that other sort of stuff. Not just Having the data in the right place at the right time, which is a lot of what they do. But from a security and access standpoint that's extremely important, especially when it comes to things like um, defense, uh, against uh, bad ah, AI actions, actors and um, that sort of thing via MCP or via prompt injections or whatever it might be. Um, so that's one of the three. The other two, um, would be uh, the platform on which you build and run and manage the lifecycle of the agents. Um, and then the third, I think, um, I just forgot actually to tell you the truth, there's a third one in there that I'll remember in a moment.

Speaker B: Yeah, when I talk about platforms, I mean platform is such a generic term. I mean when I talk about a data platform it's almost like the life cycle management around data. Right. Because in the end data drives used to be different, but today the algorithm is, is pretty generic as long as you bring the right data to the table. And so lifecycle management is really important. Um, and I look at a data platform as the component that makes sure that your data is clean, fresh, uh, always up to date, um, and allows you to um, select data based on privacy and other regulations and then you

Speaker C: have the, and also allow or not allow data depending on use and application user and.

Speaker B: Yeah, right. And I think that's at least. Uh, and again I'm not necessarily an infrastructure person from the ground up. You know, I started as a data scientist. So me a data platform is the old storage market, right, where people were talking about storage devices. To me today I don't talk about storage devices, I talk about, talk about data platforms. So it's a lot more than just that. And then you have the execution platforms, right, the super glues if you wish, like MCP and those platforms. Apple has their own platform and it's almost like we have a standard with MCP and then those platforms kind of use those MCP servers or concepts to kind of build their own world. Um, the only caveat I have with so many organizations building their own platform is that innovation goes so fast that it's going to be very, very difficult for people to stick with a particular platform. Some platforms are going to disappear and new platforms will be created. I just ask myself from a consumer standpoint, what does that mean for me if everybody has their own platform?

Speaker A: Yeah, that's true. Because in the cloud infrastructure space, um, we've seen very much that there was a proliferation of platforms and now everything is sort of coalesced around Kubernetes for example, to run cloud native applications. Um, and I think that a lot of the reason, one of the reasons that everything runs on kubernetes is not because it's the best thing ever, but just because it's a thing that can run anything. And I wonder as well, do we need that? Are we going to have that? Um, my suspicion is that companies like OpenAI, Ah, Microsoft especially, are going to be trying to position themselves as sort of the arbiters of those future um, agentic platforms. And I wonder uh, to what extent that will happen. I mean OpenAI has made a big bet to become the first mover to provide sort of the, you know, to be the Windows, um, of AI at the same time. All these enterprise companies would love to do that too. I mean I'm sure that Salesforce and ServiceNow and you know, know companies like that would love to be the, the standard platform that you run enterprise AI applications on. And I, I don't think they would even argue with me that that was their goal. So um, do you think that we will have different platforms for personal versus enterprise? Do you think that there's um, do you think there is going to be a Windows of AI?

Speaker C: No, no way. This is worse than the cloud. I mean the cloud allowed, granted there's been this force of gravity towards Linux, cloud, native and all that sort of thing. Got it. But that has not made Windows obsolete in cloud native or in web application or in application development, not to mention other operating systems. What I mean by it's worse than the cloud is the cloud birthed the API and the so called API economy, which is, you know, marketing term of art, but refers to um, APIs everywhere, all the time, such that you can plug things together. Now does that, um, does that, you know, is that it's not magical, it's not magic pixie dust. Um, you need to do a fair amount of infrastructure work among other things in order to get things to work like you expect them to. But this gets. The AI is so able to permeate every layer that um, trying to be some kind of standard in any way for an AI stack is a fool's game. It's ridiculous.

Speaker B: I think Windows, or at least Microsoft in general was trying to dominate the household, meaning that they wanted to run on the box in your house. Um, I think agentic AI platforms, I don't think they have that goal. I think their goal is to give you an API key and that you're hitting their data centers, right, because they can provide that service a Lot better. Because imagine that we all expect fast returns. We always expect that when you give a, uh, prompt that you get immediate results. If they would aim for something like Windows or, uh, a box at home, they have no control over performance and latency. So I think platforms, agentic AI platforms are not like Windows, but more like remote services where you just use an API key to hit that particular service center and then get a fast response.

Speaker A: So as we're nearing the end of our episode here, on the end of our season, um, I want to ask, um, maybe a difficult question to the two of you. And that difficult question is we've been talking about agentic AI. We started trying to define it, we've talked around it a lot. We've, we've given a lot of examples and a lot of descriptions of where companies are going, where people are, people are going. But is this really a thing? Uh, you know, let's, let's, let's meet our applicant or our audience where they're, where they are and say, is agentic AI really a thing? Are we going to be talking about this in a year or in five years, or in 10 years? Or is just this just 2025's thing and the 2025 way of talking about AI. So, um, guy, uh, what do you think? Is agentic AI really a thing with legs?

Speaker C: I think it is. I think we will be talking about it in a year. I'm not sure about three years though. This, uh, development, the development of this particular wave of revolutionary tech is so fast. I used to say about cloud computing that we would stop calling cloud computing eventually. We just call it computing. I was kind of wrong because we just kind of call everything cloud now, more or less. But it did sort of disappear as a distinctive term. And I think that agentic AI will disappear as a distinctive term and it will just be AI. Um, on the third hand, we have AGI coming, which I do not expect to be what it purports to be. Um, but I think that we're just going to start calling it all AI, including things that we don't call AI now. So agentic, um, in a year, yes. In three years, not so sure. Yeah.

Speaker B: To me, first of all, agentic AI to me, uh, is a reference to the ability to use multiple large language models, interchange daisy chain, and bring in your own data in an efficient way. That to me is the gentic AI. And so will agentic AI as a term still exist in three years? Almost guaranteed? No, um, because the marketing people will come up with Something else. Uh, but technology wise, I think the question for me in the future will be, will large language models as they exist today still be at the core, um, of modern AI, whatever it is, in a few years? Or will there be something else than a large language model? That, to me, is kind of, you know, what will happen in the next couple of years.

Speaker C: But hang on just one minute, because the second half of your question, Stephen, was, will we be talking about agentic AI in a year? Are we at the beginning or are we at the middle? As recently as agentic AI came into the discussion, are we at the beginning or are we at the middle of it? If we're the middle of it, we might not be talking about agentic AI now. Maybe we'll be talking about recursive AI, which is, um, something relatively new to me, which is AI that, that, uh, well, Frederick knows what it is. It's. It's AI that creates itself. Creates other, you know, so to speak, creates itself. Uh, maybe that's so. Maybe that's so maybe we will. Agentic AI will be so. Last year, so five minutes ago, in a year's time.

Speaker A: I think that the. If you were asking me there, Guy, I think that, uh, um, the analogy you gave of cloud computing is true. And that's sort of. I remember having that conversation 20 years ago. Uh, this isn't cloud computing. This is just computing. This is just how things should be run. And I think that most of the concepts have already been incorporated into everyday applications, and basically what we call modern applications run on modern platforms. And I, And I think the same is going to be true of agentic AI. Frankly, I don't think that we're going to be talking about agentic AI as sort of a capitalized proper noun. I think it's going to be AI agents that operate on our behalf. And I think that that's what people wanted AI to do anyway, and so that's going to be the future of it. So we shall see. Um, that's that for this season of utilizing, uh, tech. Um, thank you, Guy. Uh, thank you. Of course, Frederick, this is your nth season here supporting us on our podcast. It's always wonderful to have you. You know, Guy, it's been welcome, great, uh, to welcome you. Before we go, uh, tell us where we can continue the conversation with you, Guy.

Speaker C: Well, you can find me@futurumgroup.com. you can find my writings there. Um, LinkedIn is a great way to see, like, know when I'm engaged in a conference or a tech Field day or something else. You'll see me there. And I'm also at Guy Currier, bsky.socialluesky.

Speaker A: yep.

Speaker B: And you can find me on LinkedIn and on our website, uh, hyphens.com.

Speaker A: and as for me, you'll find me at as fosket on most social media networks. I uh, do a lot on LinkedIn. Uh, I'm on Bluesky and Mastodon even. Uh, and you would love to connect with you there. So thank you very much everyone for listening. Again, as a reminder, uh, go to Textron AI where you will find a new podcast called Utilizing AI. It's not the same format, um, but you will see these faces there. I guarantee. We're going to have Guy and Frederick join us on uh, Utilizing AI. Uh, we're recording that and uh, publishing a new episode every Wednesday. So find that uh, again at TechStrong AI uh, or on YouTube or in your favorite podcast application. Utilizing, um, tech will return. Uh, we will return with a new topic. Um, we have so far uh, talked about a lot about AI. We've talked about data infrastructure. We've talked about Edge, we talked about uh, even some, some hardcore tech, uh, CXL for one season. Uh, again you'll find Utilizing Tech in your favorite podcast applications as well. Um, also on YouTube. Uh, we, we would still love to hear from you. If you enjoyed this discussion, please do reach out. Maybe a suggestion for what we should cover next season. I'd love to be, love, uh, to entertain that and maybe we can welcome you as a guest on the podcast. Uh, this podcast is brought to you by Tech Field Day, which is part of the Futurum Group. For show notes and more episodes, head over to our dedicated website utilizingtech.com or find us on X, Twitter, BlueSky and Mastodon. UtilizingTech. Thanks for listening and we will share catch you on the next season of Utilizing Tech.