09×01 – Utilizing Agentic AI with Frederic Van Haren, Guy Currier, and Stephen Foskett

AI is the hottest topic in tech right now, evolving dramatically over the previous eight seasons of this podcast. We're kicking off season nine of utilizing Tech with a discussion of the state of the art of Ag agentic AI with Frederick Van Herrin, guy Courier, and myself, Steven Foskett. Learn about a Ag agentic ai and learn about season nine of utilizing tech in this episode.

Welcome to Utilizing Tech, the podcast about emerging technology from Tech Field Day part of the Futurum Group. This brand new season focuses on practical applications for AI and specifically agentic AI and related technologies. I'm your host, Stephen Foskett, organizer of the Tech Field Day event series, including our AI Field Day event.

And joining me for this season is a familiar face and a new one. Before we begin, let's go ahead and meet them. Well, thanks for having me.

Um, I'm Frederick Van Herrin, the founder of Hyphens. Uh, we are a consultancy and services organization helping customers accelerate their AI journey. And you can find me on LinkedIn as Frederick v Herrin.

Yeah, I'm Guy Coer. I'm, uh, an analyst at RUM Group. I'm also the chief analyst for another futurum group subsidiary Visible Impact.

And, uh, we help vendors, uh, articulate and, and, and bring to market, um, their, uh, offerings, including AI offerings. But I also have a background in market research and product management, product marketing, uh, including ai. Back before it was ai.

And I'm, uh, Steven Foskett, as I mentioned. Uh, this is in fact, uh, the ninth season of utilizing tech, uh, of those nine seasons by my Count six focused on, uh, and not including this one. I have the eight seasons previous six focused on AI and various sorts.

Uh, last season we talked about AI at the Edge. Before that, we talked about AI data infrastructure. Um, and, and as guy said, we actually started Frederick, you and I, uh, talking about AI before, uh, chat GPT was released.

In fact, we finished our first three seasons before chat, before AI became the topic that it is today. I mean, it's safe to say that as far as technology goes, AI is the most important thing in the world. Um, that sounded like a a one of those, uh, uh, movie openings, right?

AI is the most important thing In the world. Uh, do you concur? Is AI the biggest topic?

I don't wanna say the most important, but the biggest topic, Frederick? I I think so. I think a lot of the innovation, uh, that is happening today is, is heavily focused on ai, maybe a little bit too much sometime.

That's why people sometimes are kind of worried that AI is maybe a little bit too much hype as opposed to practical. But I definitely believe that a lot of the funding and a lot of the innovation today is going towards that direction. And if you, you know, you and I, we have talked so long about ai.

We have the seen the traditional ai, we have seen the generative ai, and now it's agen ai. I mean, to a certain degree, what's in a word, right? We can, we can define a little bit about Agen ai, but I definitely believe that AI is really gonna stay a, a hot topic.

Uh, the problem of course is AI is a generic word, right? So we, we kind of as podcasters, this kind of our, our responsibility to, to kind of narrow down and, and, and define things. Yeah, I think that, uh, actually your first choice of words, Stephen, were the right ones important, important in the sense, maybe not of market size or of current impact, although everyone seems to have encountered it at this point.

Um, I use the word everyone loosely, but in terms of, uh, its ability to transform for the good and for the bad, to make things better, to make things worse, and to do that, in either case, extremely rapidly, uh, I don't think we've ever seen anything like it. So I, I think important is, is actually the right word. Even if, uh, we rightly should put it in its place, explain what it is and what it isn't.

There's a lot of misconceptions about what it is and all of that. I don't think there's any more, uh, effective conversation to have right now, just pretty much across the technology and business landscape than the AI conversation. So, if that's not important, I don't know what is, You know, my litmus test for the importance of things is, uh, and, and no offense to grandmothers here, but, uh, you know, uh, have, have the grandmothers of the world heard about it, heard.

And that is certainly the case. Uh, well, my grandfather is not with us anymore, but my mother-in-law asked me about AI and chat GPT the other day. Uh, my father has said, sounds like this AI thing is gonna be replacing jobs.

Um, you know, everyone I talk to, if they find out that I'm in tech, you know, they, they wanna know what does this really mean? And I am always sounding a cautious note for them. You know, I don't think that AI is not important far from it, but I feel like at this point, we are still in the new toy phase of AI rather than the, let's get some work done here.

Phase, um, you know, Frederick, uh, you know, what do you, what do you think, uh, what I what would you say if a non-tech person came to you and said, you know, what is this, what is this ai, what is this AG agentic ai? Maybe they've heard of AG Agentic. What does that mean?

Right. So, first of all, I mean, when there, I guess there are two questions. There is, when, when people ask me about AI in general, you know, I, I try to explain it as it augments our capabilities.

I mean, you bring, you brought up your grandfather. My mother is 90 years old, and she uses chat, GPT, and I didn't teach her chat. GP gt, she's using it for translation and for writing, you know, documents.

So I think we're, we're entering a phase where when I explain AI to somebody, there's a low hanging fruit, right? It's the, the, the, the grammar, the translation, looking up things, you know, instead of Googling now nowadays it's chat GP gt. So that's, that's a generic term as far as agen ai.

Uh, the way I explain it to people is, first of all, if they're familiar with chat GPT, then I will refer to it, you know, this is a type of generative ai. And I will say that agent AI does two things, and one, the first thing it does is it introduces the concept of an agent. And an agent is like translation or sending an email.

And then the second component that makes agent ai, agent AI is the reasoning. And so the way I explain that to people is, is that agent AI is, is similar to, um, kind of thinking before saying something, right? The traditional large language models, generative AI spits out the first thing that comes to mind and, and sends it to the users.

Agentic AI is where there's a little bit more reasoning just like us humans. So in a nutshell, agen ai, the concept of agents or plugins, if you wish. And then the second piece is the fact that there is more reasoning going on than traditional generative AI Reasoning.

It's an interesting choice of words. Uh, the usual word that I stress when people ask me about AI and how to use it is simulation. AI is not intelligent despite the name.

It's a simulation of intelligence, generative AI in particular. And I, I do mention that mostly I'm talking about generative ai since that's anywhere between 90 and 99% of the attention right now. Anyway, generative AI in particular is designed to simulate reasoning or simulate, um, speech simulate.

Well, it can simulate a lot of strings, it can add a lot of strings to a lot of other strings. So it can simulate, uh, a DNA strand, for example, based on input. And I think that's a really important distinction to make.

It's the source of hallucinations, it's the source of, um, AI's general stupidity. But what AI does do in, for, in terms of simulation, is extremely useful and helpful, especially as long as you keep that in mind. So, I dunno if I'd use the word reasoning for agentic ai.

Um, I mean, the idea of agency is just that, uh, like you said, it's something that can go do things. And, um, an AI agent that can go and do things without, um, having specific algorithms or sets of instructions, um, that can more or less with permission prompt itself to send that email based on certain conditions. And then really importantly, uh, do what amounts to learning or retraining as it goes, so it can do it better.

I think that's where I land on in terms of agentic ai. Yeah, definitely. I mean, we, we can call it whatever we want, right?

Reasoning simulation, uh, or o other terminology. The bottom line is, is, is that there is data, there is, there is, there is, um, background information and historical data. And that historical data is being manipulated by math, right?

Um, the reason why in a lot of the technical industry, the word reasoning is being used. It's because it's referring to the fact that it's, that the first answer the system comes up with is not necessarily the right answer, right? So you can, you can call it simulation or iterative approach if you want.

The idea is, is that just like with us humans, is that the first answer is not necessarily the right answer. It could be, but it doesn't necessarily mean it's the right thing from a technology standpoint. It basically means there is a lot more going on when you ask the system an agentic AI system, a question, while you could ask a generative AI system exactly the same question, the agentic AI will do a lot more in the background than a generative ai.

And it's, it's very difficult to explain it to people, right? So be because people even does, don't necessarily understand to work reasoning or simulation, right? It's in the end, it's still a machine, it's not a human, right?

And nobody's trying to say that generative, generative ai, or I should say agent AI is replacement for a human, right? Yeah. And, and I think that that's the, the key there is that, um, well, I don't wanna get too philosoph philosophical here on episode one, but I do think that you could make a, an argument could be made that at some point, it doesn't matter whether it's thinking or not, if the result is the result that a thinking machine would come up with.

I think that also, it is very, very true to say that it is not thinking the way that we would consider thinking it is statistical, but that the combination of, uh, iteration as Frederick said, and, um, selective use of data can result in something that is the same effect as an intelligent system, even if it's not one of us. Yeah. And, and Agen does take us a little bit out of the bind that generative leads us into in terms of that simulation idea.

Uh, the way I usually put it is that the design point for generative AI is a simulation. It's, it's not truth, because this is, you know, commonly well known within ai, and I think a lot of the general public is picking up on this, that, uh, the, the results produced by AI can be just dead wrong. Um, the problem is that because simulation is the design point, it's appears true, it appears equally true.

Um, and, and to Frederick's point, and to your point, um, when you're thinking about agent ai, you're thinking about a process. You're thinking about, um, some automated or semi-automated process, even if, even if it looks just like the same chat bot that generative AI is, is is behind, um, that process is designed, it's designed by humans, it can be adapted to some degree by the agent itself. Um, and so the result is that, uh, you, you're, you're missing some of that simulative character.

So I don't think that's philosoph philosophizing at all. Um, I think, um, it, it's a useful corrective for us to understand what's going on, uh, right now with AI and what it's, what its promise is. I just worry that even the creators of these, um, agents, um, are fooled themselves into, uh, into what they're capable of and what they're doing.

Well, hopefully that won't be happening here. Um, I think that, uh, we've got some, some folks here who really do understand, you know, what's really going on. Um, but you know, Frederick mentioned another aspect as well of ag agentic ai.

That's, I think, equally important, and that is the ability to, in a way, to perform work. And of course it has to have, uh, context, it has to have a chain of thought. Uh, it sort of an iterative re reasoning process to analyze that data and decide, you know, not to, I, I'm anthropomorphizing to, uh, Output A, an action, and then it has to have the ability to take that action.

And that has led to a need for standard frameworks to allow these AI agents to interact with each other. And one of the ones that we're hearing a lot about, and I think we're gonna hear a lot about this season, is what's called MCP or Model Context Protocol. Um, which of you would like to explain what a, what MCP is?

Well, Frederick's been on the firing line first, so I'll go first, then he will correct me if that's okay. Um, because I think of it as pretty relatively simple. MCP is a way for, um, AI based applications or AI agents, uh, to seek context, um, to request context, and to receive it.

Um, largely it can be from other you knowis or AI engines. Um, and it can do this using a relatively standard a p iLike interface. So it can be programmed in, or it can be, it can, uh, uh, discover these, uh, resources, you know, in its system.

And that is what allows these, uh, systems of AI, including AG Agent AI to be more effective and to, and to work together. Yeah, exactly. It's, uh, it's, you know, agenda AI is, as I mentioned just about agents.

You have different agents. A lot of organizations are deploying and delivering agents that consumers can pull together. And an MCP server has the ability to pull all these agents together and generate the content.

I mean, it's, it's important to note that there's, there's a few versions of the MCP server. You know, some are task driven, others have a, a different, a different approach. But in the end, you can look at it as a, as a way to standardize, right?

You, you, you have a bunch of agents that are very capable. Um, you, you have the ability to daisy chain those agents, right? And so you can build very, and when I say you, I mean, you as a non-technical person or consumer can build a reasonable, workable, uh, application with MCP servers.

It has to be said that MCP, there are probably like two or three different server types today. Um, it's evolving really quickly, but you can see how many organizations are jumping on board and delivering capabilities, right? So for example, Docker desktop is, uh, is an application install on your desktop, which comes with, with MCP servers ready to go.

Uh, and can Steven, can I add a little, a little con uh, not context a little bit to, to this, the importance of MCP course MCP being an open standard. So, you know, uh, to just give a general label to all this stuff, ais have interacted with each other before, programmatically before, uh, less than a year ago is when the first MCP specification was published. I mean, this thing has grown super rapidly.

Um, but here's what I wanted to say. The import of something like MCP cannot be overstated. If you think of just a regular generative ai, um, uh, model, it's taking a string of things and outputting a string of things that should follow that string of things.

Usually the string of things is words, and it follows with more words. So when you are doing good prompt engineering, you're adding all this context and all this stuff to make that input of words and attachments. They're all, it's all, you know, a string of things to generate more things.

The more you provide, the more complete, the more on point it all is, the better your output. That's generative ai. Now, imagine that the AI did not have to rely on whatever you happen to put in, but could go out and seek other contexts.

That's what MCP allows. That is critical for an AI to be agentic and not just generative. Yeah.

And you know, ultimately, like, as, as Frederick was saying, I think the, the thing about MCP that is exciting is to me, the way that it encapsulates this context in a way that is standardized. In fact, I could see MCP being leveraged by non gen AI technologies as well, because it is very much, it, it just makes a lot of sense. Those of us who've been using, for example, process automation technologies for, for years now, or this sort of, if this, then that type technologies have encountered the problems of, um, basically AI rot or API rot, um, uh, making sure that as things are upgraded, that they continue to work.

Um, figuring out how to pass data from, um, I hate to use the word agent, but from agent to agent, from component to component, and MCP actually, um, takes a lot of that work. And in the context of generative AI moves it forward into a extensible framework. Now that's exciting beyond ai, but in the context of ai, it's especially exciting because what it means is that you can basically give, uh, a package a payload to the next, uh, worker in the chain, the next ai a, you know, agent in the chain and say, here, do something with this.

And unlike conventional APIs that are somewhat brittle and fragile, uh, it can be a lot more robust because it uses generative ai. At least that's how it's been to me. Um, what do you think of that, Frederick?

Yeah, that's exactly right. I mean, uh, i, I know you don't like the word agent, but the agent in an agentic AI environment doesn't have to be AI driven. It can be something very, very simple.

Um, and to your point, you can have agents that are non-AI driven, but by, by enabling it with an MCP server, you end up with a, an application that can do a lot more than the components individually by himself. I'm not sure if we actually defined MCP, you know, it stands for model context protocol, uh, in case people wanna look it up. Um, but you're absolutely right.

I mean, I, I think what it's, what what AgTech does, what Gentech AI does for the community and people out there, it, it enables people to do a lot more. And we see that, right? People that we're asking for basic applications in the past are now asking for similar applications, or at least similar functionality, but then driven by an MCP server.

And it's, it's, it's fascinating how easy it is to set it up, right? And, and we have said it before, but the, the, the, the speed of innovation is incredible. Um, certainly combined with vibe coding, I mean, who needs an engineer to build a prototype?

I'm not talking about production, but prototype wise, it's, it's an incredible time to, to be around. I do think that we need engineers, and I, I know you, you're not being an absolutist about this, Frederick not at all. Um, but it's that, it's that whole idea that when you're using generative ai, for example, it really helps to have expertise in the area that you are working on, um, so that you can utilize what comes out, um, for good and not, and recognize the part that might be problematic.

And in the same way, um, I I, not for nothing. I think, uh, you know, if if there's such a thing as elegant code, there's probably such a thing as elegance in a vibe code. Well, sure.

I, I could, uh, I, I'm with you on that. I, I actually am concerned that as people are vibe coding more, they may mistake vibes for quality and think that they actually have developed not a prototype, but afin, but a finished product. That's right.

But that doesn't sound great. Um, but that being said, I hope that, uh, I hope that that won't happen. And I am actually, um, you know, fairly optimistic about a lot of the work that's been happening.

I mean, if you look at what MCP does, it constrains the context that, um, the next link in the chain can work with. You look at some of the other, um, guardrail type, uh, things, uh, that, that are being put up around, um, AI systems. I think that that's all good, because a lot of the problems that we've been having with ai, um, you know, I mean, certainly my biggest problem with using AI is that it's non-deterministic.

Um, you know, I can throw, uh, a set of data at Gemini and get this output, and then I can throw it the same set of data at Gemini and get a completely different output. And that's challenging for me as a, as a developer. I think that there's, um, many ways in which we can kind of address that with additional guardrails and boundaries and context setting that can hopefully help kind of constrain some of that randomness.

But at the same time, um, I do think that it's exciting where this stuff can go. Um, again, you know, one of the, the words that I used before was brittle. I have found, um, agentic systems prior to AI to be extremely brittle, to the point that I became very frustrated in a lot of these process automation technologies, because essentially those links in the chain would be changed without notice somewhere.

And so, even though it was deterministic, it always gave the same output. Um, it sure didn't once they changed the API on me. And, you know, I actually, in this actually, these days, I'm, I'm using, uh, generative AI as sort of an A API super glue already, and where I'll throw it some JSON from something that I know sometimes is a little bit iffy and say, give me a js ON output from this JSON input.

And, and the result is usually a lot more sturdy and reliable than, than anything else. And that's what I'm hoping that we'll see with Agen Paths agent to agent and MCP. Yeah, so we, we talked about two different sides of Agen ai.

One, one is a developer site, which is, which is an interesting piece by itself. But, but I, I have to reiterate, what I always say is, is AI in general, whatever it is, is, is to augment our capabilities not to replace. So you will never hear me say that, you know, vibe coding replaces a, a developer.

Um, when I use Vibe Coding, if I even can call it like that, it's the equivalent of me buying a book and looking up for a reference. You know, an API call now I go to Vibe Coding and, and Pro, and it'll provide me with some, some reasonable, um, guidance around API calls. And then, and then there's the flip side on, on people consuming Agen ai, right?

I mean, I, I think another, another thing I, I have a problem with, with people kind of assuming that whatever Agen AI spits out that it has to be exactly, uh, what you expect. I mean, it's, it's having different opinions. It's not bad, right?

It's the same data, different opinions. That's, that's what we all do, right? That's why we have this conversation.

We all have the same data or similar data, but we might have a different, different opinion. I think it's important to, to note that the Gen AI by itself, um, might give you different answers and, and evolves, right? I mean, another thing which we haven't talked about Gen ai, but RAG is really important, ingen ai.

So RAG is the, the retrieve, augment generate, which is the, the ability to inject almost in real time information and change the behavior of Angen AI system, right? So, so the expectation is, is that the system should behave differently if you're asked the same question over and over and over. Yeah.

I, I'm looking forward to, to learning how, um, practitioners and, and, and, you know, I suppose vendors as well are, um, uh, putting, putting borders around or, or I, I guess identifying scenarios is really what I'm really thinking of. That here's a good scenario for this type of AI work. Here's a good scenario to avoid.

And I don't mean, I, I guess the reason I eventually avoided the word scenarios is I'm talking about sort of not, uh, oh, this is really great for computer vision. This is really not that, not that kind of scenario. I mean, scenarios where the type of work, the type of environment, uh, the type of decision making required, um, some will be obvious, uh, regardless of the application for AgTech or for generative or for both, and some will be obvious ones to avoid.

I think that that kind of, everyone's throwing AI at everything all the time right now in a certain sense. And that's makes sense when no one is really sure exactly, uh, where it's going to be productive and we're not productive. But I, I, I wonder if, if there are, there are practitioners out there right now who have enough experience at this point to be able to say, no, we don't have the right personnel, or we don't, this is not useful for this particular type of work.

I'd, I'd like to find that out. So as we look forward to the next, uh, you know, uh, eight episodes of this season, I wanna take a moment here before the end to ask each of you, you know, what, what would be your ideal, uh, outcome for this? What would you like to learn?

And who would you like to talk to, uh, over the next, uh, coming weeks to learn that, to reach that, um, guy you wanna, you wanna kick us off? What would you like to learn this season? I'd like to learn if, um, twofold, if there are, um, productivity measures that are, that are, uh, lighting people on fire.

I've been maintaining from the beginning that productivity is a secondary benefit of ai, that it's just, um, it helps you or humans or certain types of work to be more reliable because you can just get started instantly. No writer's block. Um, so it's more reliable and that you can fit more review cycles in, so you can come up with higher quality work and that productivity flows from that.

But I do think that productivity is why everyone's in it, and I wanna understand, um, what people are seeing. And I think there's less productivity out there than, uh, advertised, but people are still pursuing it. So why they're doing that.

I, I think that there are lots of benefits that don't necessarily come down to dollars and cents or ROI or that sort of thing. And, uh, I, I would love to help understand and shape the discussion around that. Would you, Frederick?

Yeah, I think, I mean, the, the engineer in me says, I wanna learn about innovation, right? What do, what don't I know and what's around the corner? Um, and I think that's, that's the first thing is always to learn something from, from other people.

Um, the second thing is, is, um, the, the fact that that systems are becoming more and more complex, it's, it's to the point where the people who provide the models and provide agents don't even know how their, a final product will be used. So it's governance and security. I really would like to find out, and this is the holy grail, right?

How do you, how do you diagnose or analyze a given, uh, agent AI system for governance, security, and bias? Even, even today? It's a problem, right?

We're, we're, we're giving a system and we have no good metric to validate those components. And as technology goes faster and faster, there is a tendency for, you know, leaving that behind or as an afterthought. Um, which, you know, Steven, you and I have been talking for it for a long time, you know, governance, security, um, is, is a big concern, and I think it's, it's getting worse.

And then my final statement is I look at AI as an assistant, so I would like AI to be a better assistant to, to me in my work, uh, in my private life, Really Good points, especially the, the, you know, security or, you know, let's not use a fancy word, let's just say sort of, you know, human control if you like or something. I think that's a big worry. And maybe that's an area where we need to mature a bit.

Uh, just saying like, you did Steven, like, it's not deterministic. That's a fancy way of saying, I don't know what it's gonna do. And, and, and, you know, that can be a problem, but that's true of the humans we interact with, so, you know, better get used to it.

That's certainly true. Uh, you know, I didn't expect you to No, uh, that, that's certainly true. Um, and, and I would add, you know, one more thing I'd like to see is I'd like to hear, um, about productive uses of this technology.

I really wanna know what are people doing with this that they couldn't do before? And that to me, is the hallmark of any kind of successful technology. I think right now we've, we've, we've got a really cool thing going, but we need to make sure that this isn't just a parlor trick.

That this isn't just, um, a toy. It needs to be something that's useful. And so, again, back to the title of this podcast way back eight seasons ago, uh, when we said utilizing ai, why did we call it that?

That means to make productive use of a technology. And so let's figure out how we can actually utilize AI now that we've got technology that works now that we have a context protocol, now that we have the ability to connect AI with external data sources, uh, we've got infrastructure, um, how are we actually using this stuff? And that's actually one of the things we're gonna talk about on the very next episode.

So on, on the first episode of the regular episode of this season, uh, we're gonna be talking to, uh, a great, uh, a great leader and thinker on this about how, um, his company is building AI models and agent systems, uh, that are specific to, uh, industry verticals. So they're not just putting a chat bot on the side of the website to say, how can I help you? They're building applications that do things in specific verticals, and, and so you'll learn a lot more about that.

And over the season, we're gonna be inviting more people like that, whether it is companies that are designing and building products or, um, thinkers, uh, doers who are out there creating this or, uh, thinking about it and advising on it. And hopefully, uh, when November comes around and this season is, uh, is done, you will have learned a thing or two because I'm pretty sure that I will have, uh, along with Frederick and Guy. So thank you very much for listening.

It's great to have you join us for this season of Utilizing Tech. You will find this podcast in your favorite podcast application. You can also find videos of it on YouTube, and, uh, you can find it streaming in the Techstrong app on Roku and Apple tv and other places like that.

If you enjoyed this discussion, please uh, leave us a rating. Leave us a nice review. Uh, this podcast is brought to you by Tech Field Day, which is part of the Futurum Group.

com, or follow us on X Twitter, uh, blue sky and Mastodon at utilizing Tech. Thanks for listening, and we will see you next week.