Latest / Practical AI / Models, Harnesses, and Multi-Agent Systems

Transcript

Models, Harnesses, and Multi-Agent Systems · Practical AI

0:01
Narrator: Welcome to the Practical AI Podcast, where we break down the real world applications of artificial intelligence and how it's shaping the way we live, work, and create. Our goal is to help make AI technology practical, productive, and accessible to everyone. Whether you're a developer, business leader, or just curious about the tech behind the buzz, you're in the right place. Be sure to connect with us on LinkedIn, X, or Blue Sky to stay up to date with episode drops, behind the scenes content, and AI insights. You can learn more at practicalai.fm.
0:35
Now onto the show.
0:41
Daniel: Welcome to another episode of the Practical AI Podcast. This is Daniel Whitenack. I am CEO at Prediction Guard, and I'm joined as always by my cohost, Benson, who is a principal AI and autonomy research engineer. And on these episodes where it's just the two of us, no no guests, we like to take take a moment to really, sometimes talk about just a topic we wanna talk about and learn about, sometimes be more educational, dig into specific details, and I think that's more of the direction we wanted to go today. Right, Chris?
1:17
Chris: It is, and actually, we've done several of these recently, and I think it's important to call out. If you've been listening to some of the recent episodes, there's so much happening right now that one of the priorities for us in the show is to try to help people really stay up with what's going on, not just from a newsy In
1:37
Daniel: practical way
1:38
Chris: so that you can actually go do this yourself in your organizations. We kind of realized that while there are some We always have a huge funnel of guests to come in. We realized we needed to kind of pause and take a few episodes and try to just get people caught up to where they need to be, because so much is happening right now, and so that's a big part of why we're doing this right now, and so we hope this is gonna be helpful for a lot of folks out there.
2:08
Daniel: Yeah, This is always evolving, and the way people use terms is always evolving, so it's it's useful to even sometimes redefine some of those terms. And I I promise we won't belabor this point too long, but I do I do think it's worth us defining a few terms here up up front so we're just all on the same page. We're talking about the same thing, and some of these we've defined before in other episodes, like you said, but again, the semantic variance of terms changes over time, and it's, it's worth just updating and and reminding. So the the first term obviously is an AI model or a model. And very often, and this actually hasn't changed, although maybe some of the things we refer to as models have changed.
2:59
But an AI model is really if if you're more coming from, like, the technical side, you might think about this like a function or a software function. Maybe if you're coming from another angle, really, it's just a transformation of data. Right? Like, I give you an image and you tell me what's in the image. It transforms to objects in the image.
3:21
That's an object recognition model. Right? Or I give you a series of words and you tell me what words should come next. So that's a large language model. It's a transformation of input to output.
3:35
It's a function implemented in software, and these models generally do a a single type of transformation. So a large language model does that transformation that I mentioned. It takes input words or text and then generates text output. Language vision model changes the input side changes, so it might take language or text plus images or just images and then produces text output. An image generation model, as the name might suggest, takes in text and goes the other way.
4:11
It generates an image out out the other side of the function or data transformation. There's, you know, image basically, any combination. Just think of the model as that data transformation, and then you could have a lot of things on the input side. You could have a lot of things on the output side. And depending what type of type of data is coming in on the input side coming out on the output side, we refer to these as different types of models.
4:40
So an LLM, a language vision model, an image generation model, a video generation model, forecasting model, an autoregressive model in that case, an anomaly detection model, etcetera, etcetera. There there's all sorts of types of models, and not all these are neural network based models. A lot of the ones we typically think of today are, but a model itself could be composed of guts that take a variety of forms, and the the way that that function does the transformation, that could take a variety of forms. But at the end of the day, it's a function or a data transformation.
5:23
Chris: No. That's a great explanation. And and it really like, the architectures have gotten much more complicated over the years able to do a lot more functionally, but at the end of the day, you mentioned neural networks and stuff. It's still neural networks kind of plus plus architecturally in terms of how that function is, but I really like the notion of a function because you give it an input and you get an output, which in this case we call an inference, and kinda keeping that in your mind as to what it is is pretty important because it's not magic. We get a lot of that in mass media about AI, and so yeah, we're talking neural network based architectures that are able to do transformations along the way, and now we're building all sorts of stuff around it, like agents, you know?
6:15
Daniel: Yeah. Exactly. And and maybe one distinction too that's really important with these models is some are what's called open weight or or open source models. Some are closed models.
6:26
Chris: Good point.
6:27
Daniel: And and so these models can take at the end of the day, they're all kind of doing the same the same sort of thing, but the way that you can access them as a user might be different. And some people might have seen in the news, even just in recent days, NVIDIA and others making pleas to the community around the importance of open weight models or open source models. So what this means is if you're gonna run that function, if you're gonna run that data transformation, so you're gonna take text in and generate text, or you're gonna take text in and generate an image. Some there's really two components to that. One is you have to have software that runs that model and actually executes it.
7:12
The other thing is that you need what are called parameters or weights and biases. These are just a dataset of numbers that configures that software function to operate in the way that it operates. And so you need both of those things. And sometimes you you as a user may have act direct access to both of those things for a given model. So let's take, for example, an an open model, like a maybe a Gemma model or something like that.
7:44
You can go to Hugging Face. You can download the weights of that, and there is software that runs it, the transformers library from from Hugging Face, for example. And so you can have both of those components. You can run it, you know, on your own laptop in your cloud environment wherever. Other models are not released, either the software component or the and or the data component of the weights are not released, and those are generally called closed models often kind of just accessed through a managed API.
8:19
So programmatic interface or a chat interface or other interfaces. And those are sitting behind a company's proprietary infrastructure. They're still running the software. They they still have the same thing that they're running on their end, just their version of it, and they keep that proprietary as their IP. So that's what we mean by open weight or closed.
8:40
Chris: One more thing to throw in that, because it's relevant to the rest of the conversation, is these models can be of many different sizes, What we call frontier models, which are the largest, most complicated, and produce the most sophisticated output often require hardware that is going to be prohibitively expensive, and so that is a huge investment if you're gonna go that way. The cloud providers and these top tier model creators have enormous data centers filled with server racks to manage this. Not all models are at that level. You mentioned Gemma, which is some of the gradually smaller models that can be downloaded and run locally with whatever hardware you have available. If they're open, well, Gemas are open, and there's others as well, both out of Europe.
9:38
There's a lot out of China, which is kind of the palette house of open models Can
9:43
Daniel: be. Yep.
9:44
Chris: And And so, it's interesting that we're in a moment where if you're wanting to run a fairly sophisticated model on your own, on your own server, you're probably most likely going to a Chinese model these days.
9:59
Daniel: They're the best.
10:01
Chris: They're the best right now. Or if you're willing to do APIs for closed models, the American based ones are still leading in that space. So that's just a general shakeout of where things are at this moment.
10:14
Daniel: Yeah, yeah. Okay, so model, open weight, closed. What the agent, and or the agent harness, Chris? What are we talking about there?
10:27
Chris: Yeah, we're like, if you just, we've used this before. If you think about the model, that function that's trying to do something fairly sophisticated, its output is kind of being the brain, then historically we've had these chatbots, and there wasn't a lot. It was more of the Oracle was sitting there and you gave it some input and it gave you some output. It was still like, Oh, okay, I'm having a conversation. But the big change in the last year has been AgenTx, and that's where you have one or more agents that is a type of software that is a task to go do something out in the world.
11:07
It might be on your system. It might be in robotics. It could be almost anything that you want it to be, and you can have one or more that are working together, and they are managed by that agentic harness that you referenced, which is a software system. I've heard it described as an operating system for AI, where it is accessing the models that you're using. It is managing the agents and orchestrating what they are doing and how they're working on that, and so it kind of gives you a body of software that you're able to use to get productivity done, to get productive tasks done.
11:46
So agents are smaller bits of that that are in the harness, but it's all just software, including the model at the end of the day, and that's important to keep. Some of these can require sophisticated hardware, but at the end of the day, the model, the agent, the harness, all these are just different types of software that work together.
12:07
Daniel: Yeah. May maybe one way to think about it is you have that one model, which is a function, an input output, and the amount of code and software surrounding that to do various things can vary. Right? So initially in the very early or or in the earlier days, and earlier days, I'm meaning, you know, a couple of years ago or or whenever it was, we we had software that was basically just a chat interface, a thin chat interface. And to your point, there wasn't a lot of business logic between that chat interface and the model.
12:45
Basically, what you put into the chat interface, that was almost exactly what went into the model and what came out of the model was almost exactly what came back to you. Now the amount of business logic that you can apply on top of that model can vary widely. Right? And the but that would be kind of like AI embedded in your software. Think the distinction of the agent, which is very interesting is and what I would consider an AI application or an AI feature within your software is really something that is meant to have a one to one interaction with you as a user of that software.
13:22
So a chat interface, go in and I type something in and I expect a response back, and then I type something in. It's turn based. Or maybe there's things like in financial software where you go in and there's an AI model that makes a prediction for a future, you know, cash flow, variance or or something like that. Right? And so those are AI features.
13:47
What I think is different about the agent, again, it to your point, is not fundamentally different in terms of how you it's still software that surrounds the model. Right? But it the what the software is meant to do is to accomplish a goal and in many cases to operate with autonomy. And that's where I think the agent is different. So it's tied into systems that are within your company or within your context.
14:17
So if it's a personal agent, maybe it's tied into your email. It's tied into your WhatsApp. It's tied into your Google Drive. It's tied into your messaging interfaces. It's tied into your calendar.
14:28
Right? If it's a company agent, maybe it's tied into your NetSuite, and it's tied into your Workday, and it's tied into your transactional databases, etcetera. And you express an outcome to that agent and or encode a goal with that agent, and it operates at least with some level of autonomy without a back and forth with a human to accomplish that outcome or maintain that goal or accomplish that goal. I I think that's that's really a a key differentiator when it comes to an AI application or an AI feature versus an agent.
15:09
Sponsor: It's more important than ever for businesses to move super fast with updates to their website, to their landing pages, to their SEO, to their AEO content that is surfacing things and answer engines.
15:25
And it's just not sustainable when a new landing page or website update turns into a pile of tickets and handoffs. Our our partner Framer actually makes that process move so much faster. Teams can collaborate in real time, iterate on the same pages, and publish instantly. Thousands of businesses already use Framer as a pro website builder, I'd really encourage you to check them out. You can learn more and get more out of your site from a Framer specialist or get started building free today at framer.com/practicalai for 30% off a Framer Pro annual plan. That's framer.com/practicalai for 30% off. Framer.com/practicalai. Rules and restrictions may apply.
16:18
Daniel: Well, Chris, the the that was, I think, a great a great set of hopefully hopefully, we didn't muddy the water within any within any of our definitions of of things, but I think that was that was some good context setting when it comes to some of the terms that we're about to to use. Did did we miss anything, and what what is your thought?
16:41
What what kind kinds of questions have been on your mind recently in terms of clarification?
16:46
Chris: Yeah. I think you set the stage really well, and I think in terms of that connection that the agent has to kind of real world things that a person's company or it could be in their house. It doesn't have to be a business. There's all sorts of different contexts here, but you're giving those agents access to things that matter is the bottom line, and you're giving them tasks to do on how to do that. And you were kind of addressing the single agent moment, but the next level above that is to start having multiple agents that each have purposes to a task level, and then they are starting to interact.
17:26
And when you get to that level, you really start That's where the autonomy of what agents are, or the potential of what agents can do is really realized, because instead of you as a human kind of being the recipient of each thing there that the agent is doing, and it's really almost an assistant to you, you can actually assign agents to go out and work together to accomplish multiple tasks in a complex environment, and where different agents have different purposes and expertise. And so they are gonna be able to kind of focus with that, and then after that, they're going to be able to, interact with each other and get this work done.
18:13
Sponsor: If you're listening to the Practical AI podcast, I'm guessing that you value practicality, not just the hype around AI, which is why I think you should check out the Midwest AI Summit. This is an amazing event. I'm gonna be there this year. It's happening October 15 in Indianapolis. This is an event like no other I've been to.
18:35
There's actually an AI engineering lounge where for free, you can go up and get expert advice from practitioners and get feedback on your architecture, your design, your agentic harness. Whatever you're looking at, you can get feedback on in real time in between amazing speakers that are on the main stage. So don't miss this event, again, October 15 in Indianapolis, and you can use the code Practical AI 20 for 20% off. So go to midwestaisummit.com and grab your ticket today. Use code Practical AI 20 for 20% off.
19:13
Midwestaisummit.com.
19:15
Chris: So that was a great explanation of getting us kind of, you know, from basic models into the world of human model of chatbot and into the world of agents and what's different there if you introduce the agent into the process. After that, you can start multiplying out these resources. So by virtue of having that agent, you have the harness that the agent is working in. It might be fairly trivial if you're just barely getting into it, but you can actually start expanding those and looking at what different types of more complicated tasks you want to address where you have not just one agent, but you might have multiple agents that are each addressing different specialties, if you will, and potentially each of those agents has access to different resources within maybe a larger enterprise, where not one agent can go everywhere and do everything, but they have to start interacting to get job done. And you mentioned prior to the break that as you get agents, you start moving into some level of autonomy and permissions in terms of what they can do.
20:28
And so as you get multiple agents and you start deciding what agent roles are going to be and how those agents should interact and what they have access to, then you can really start getting a lot of productive work done through your group of agents at this point that is largely autonomous, and you can give the system instructions that have specific outcomes that you're looking for, and you either encode those into that, and they're gonna go and work within the parameters, as we've learned with some of the security breaches lately, within the parameters that you set, assuming that you set them effectively enough at this point or you have agents doing that for you, then they're gonna go do a certain number of tasks that you've identified as productive with specific outcomes tied to those. And you can get lots of really productive work done this way, and this is kind of where things are right now. The one other thing I'll throw in before throwing it back over to you, Daniel, is that along with multiple agents, people are starting to recognize that different types of models can contribute. You can have larger models for some things, smaller models for others.
21:48
Some are more specialized, some are more general, and depending on what you need to get done, you can start having an architecture with multiple models and many agents doing this. What's your experience so far in this realm?
22:01
Daniel: Well, I might have a follow-up question for you because I know you obviously spend a lot of time thinking about autonomy and even you know, fleets or or swarms of of agents or, autonomous entities. But, I I'm wondering, people may be very familiar with the, you know, some listeners may have used AI systems a lot where there there was that turn based thing, like, oh, I'm gonna I'm gonna always have my chat GBT tab up or my Claude tab or, you know, Claude code, and there's a back and forth going on with with those things. What could you do you have any examples in your mind of things that might be impossible in that mode for, like, a single agent maybe interacting with a single human, but that would be possible for a fleet or a swarm of agents that's operating, you know, more autonomously in the background. Do you have any examples that come to mind or or things that you can think of that that people could grasp onto concretely?
23:11
Chris: Yeah. I'll try to do that. A little tricky because of the the space that I work in, but I'm gonna try to genericize this a little bit. You mentioned the notion of fleets and swarms of agents, and those are actually two different things. A swarm is usually a fleet, but a fleet is not necessarily a swarm.
23:30
And I don't want to get us off into an autonomy conversation about what this
23:33
Daniel: There's a changelog episode about this, so maybe we'll link that in the show notes.
23:40
Chris: We could do that. So when you have a lot of agents, and you're giving them each certain capabilities and access to certain things, then you can start thinking about situations that you might have been afraid to use systems in before and places where you definitely want a person controlling everything because of the time sensitivity or care that they have to bring to bear. And so we talked about cybersecurity in last week's episode, and that's a good place where if you're under an attack and your organization is needing to defend, it might very well be there are many agents, as we saw in the last episode, that were involved, but your defense against that as well might be many agents that are operating either as a bunch of individuals with some crossover or a swarm where there's a lot of crossover communication and task sharing between them, because you're humans that suddenly your systems are under attack. Maybe it starts with a denial of service or something that you notice, but then you notice that there are breaches in different systems, and it's happening very rapidly in 2026 compared to even 2025. And so at this point, it's very difficult for the humans to keep up with that, as we talked about last week.
25:06
And so you can have agents assigned to operate your cybersecurity defenses in different aspects of your organization and each one to be responsible for different parts of that so that as breaches occur, you have different agents responsible for that. So that is kind of a broad use case because everyone needs some level of cyber in today's world, and if you're not addressing cyber, you're gonna feel pain at some point. So that's something that goes across industries. Another quick example, without diving too far into it, is that though most people are not used to pervasive robotics everywhere now, at least in most Western countries they're not, we're right on the cusp of that, and as production capabilities are getting less and less expensive to do it, and capabilities of what different types of robots and drones can do, we're seeing more of that. And I mean, case in point, Walmart advertises drone delivery, as does Amazon, and we're seeing more of those.
26:10
You'll start to see robots popping up around. I know you have already done that near your house. I remember talking about that, and this will happen more and more. And agents in robotics are incredibly useful because there's a lot of things that have to happen, both from a functional standpoint, from what I would call a mission standpoint, but you could call it an objective standpoint, and from a safety standpoint in terms of not accidentally doing harm out there in the world, which is easy because you have a physical device that's doing something and interacting with real world things, including people, and so there's a whole set of considerations. And having agents that are designed to handle those tasks and having the resources on board so that they don't lose contact, they don't have to worry about losing contact with a cloud provider or something, is becoming increasingly important, and we'll see more and more of that.
26:59
So these are all One is kind of out in the world with us, and one is cyber, but there's a lot of places that I'll bet people can find in their own industry where the necessities may be real time or near real time or just such complexity that having lots of agents interoperating together is gonna become necessary.
27:19
Daniel: Yeah. And I think you can start to think through some of those flows in your own business as well. Like, if you think in manufacturing, for example, you may get notifications from some of your suppliers that then create a trigger for a first agent to actually process that notification and realize, oh, we've got a vendor problem here. And then maybe there's another agent that does research for, you know, vendors to replace your raw materials that are now at risk within your supply chain, and that's automatically triggered and just happens. And then that generates something that is then output into orders or or draft information that goes into NetSuite and interacts with that system.
28:10
And all of these things could be happening in the background. One of the things that I'm experimenting with, Chris, and I forget if I our listeners will will forgive me if I'm repeating an example that I already used, but I've been experimenting with this, with this example in workshops and discussions that I've been having. Over the last over the last few years, I've got into to motor racing quite a bit, not myself, doing any of it, but just as a fan. And I remember always throughout my career, everyone would use the example of the f one pit stop as in kind of like leadership discussions and team discussions about how a team should work together well. Right?
28:53
And and often how that would go is, like, you'd look at how the pit stop evolved over time from, like, whatever, twenty seconds down to two seconds. And then you realize, well, the the way that they're able to do that is they literally have a single person do a very specific task and nothing else, like move the tire from here to here. And that's all that's all they do, and everybody coordinates well with one another. And it used to be in those leadership or team building discussions, they would talk about, you know, how the team could work together well when it was people. Now I think what's interesting is every individual human working in a business, like me as a CEO or you as a research engineer, whoever that is, we're actually elevated from doing a a single task really well in our position to we're actually elevated more to the team principal or strategist level of the f one team.
29:52
And now each person can actually have a high performing team of agents under them, each of which do various tasks very well. And you're now thinking more at the strategic or outcome based level like that team principle, like that strategist. Right? So I think actually if you think about it this way, it does a couple things. One is it it kind of empowers people to maybe think about all of those elements in their job where they might have this team of of agents to to help them and actually operate on their behalf.
30:27
It helps you understand maybe, like, the level at which you should be thinking, and I think it actually elevates the human in the situation because now they're able to actually apply their domain knowledge, their strategy, their their kind of outcome based thinking, towards towards the goals that matter for the for the business. Now that's the more positive side of things. Obviously, you could argue, well, not everybody's gonna be operating at that strategy outcome level. Then how many strategists do you need, and do you need less humans, and what is everybody's job? And there's a whole series of questions that could come from that.
31:06
Right? But I find it, at least to be a useful, you know, a a useful tool to think about how this might operate in terms of your immediate surroundings and your business.
31:20
Chris: That's right. And it's important to recognize that this is you know, one of the reasons, like I said at the beginning of the episode, we're doing some of these episodes, addressing some of these are this is happening really fast in terms of the availability, but so is the understanding that we really wanna bring all people along so everyone can benefit from this capability. But this is changing global economics gradually, and so it's not something It can seem very esoteric to a lot of people who aren't familiar with where AgenTicks are today. If I sit down with my extended family at a family outing and try to have that, and they're aware that something's happening, but they don't know it super well, and so we'll talk a little bit about that, but this is the kind of thing you really want to get on top of, because it's moving quickly and it's creating huge competitive advantage in some organizations, while others it's diminishing rapidly. So I just wanted to draw the importance of kind of grabbing what the agentic economy looks like as early as possible.
32:30
Sponsor: As you're listening to this episode, you can tell that there's a lot to think about. There's a lot to know when you're scaling up your digital workforce of agents. I'm leading a company called Prediction Guard, which provides an AI control plane to help you actually scale your digital workforce in a governed, controlled way. We help you look at your supply chain, including models, MCP servers, making sure that you know what's in your supply chain, that you can track that. We institute runtime governance over all agent interactions, controlling that behavior as agents operating in your environment, and we provide observability around what those agents are doing and any policies that they're violating so you can be sure that your agents are under control and being scaled in a way that's governed. Please check us out at predictionguard.com and book a call with myself and my team this week to talk through how you can scale your digital workforce while applying zero trust and maintaining control. That's predictionguard.com.
33:37
Daniel: Well, Chris, I I, of course, you know, it's probably no secret to our listeners. We're always using AI to help us do these interviews or do research or or whatever. One of the things that that I did in prepping for this conversation, Chris, was just get get a few prompts and maybe questions that people are wrestling with right now for for us to consider.
34:03
And so I don't know how many of these we'll get through, but I thought that they were interesting prompts to to us. Okay. Which, one one of them which I thought was interesting is is multi agent architecture, so these fleet swarms, whatever you consider, is that really the future, or are we just compensating for limitations that better models will remove? Meaning, models seemingly will get better. So do we really need multi agent systems, or is the better model just gonna be able to solve those problems?
34:39
Chris: So I think it's an apples and oranges thing, and that is, we've talked about A third route. Yeah, there you go. I think will be Models will get better, and there may be some models that can do quite a bit for you without deploying multiple models in your system. Depends on the nature of the system, obviously. There will be other systems, especially as you're moving away from where you can host giant frontier models of the future to where your resources are more limited, where I think you're going to be dealing with lots of smaller models that are more specific on tasks.
35:18
And at the end of the day though, you're still talking about those are functions. At the beginning of the episode, you gave us a great way of understanding what models are, and at the end of the day, those are functions, and they gotta do stuff. And so I personally am quite convinced that multi agent architectures not only are where things are going, I think they will be absolutely and completely pervasive in the future, to the point where we won't wanna talk about them at all because they'll just be so built into the fabric. And so I'm actually actively writing a book on this topic, among other things. So yes, we are going into a world where this is not going away.
36:00
I'll leave it there.
36:00
Daniel: Yeah. I I think that's a great perspective. And another prompt here, well, another one I I found interesting is, do we think we're moving towards a future of more interchangeable models in these agents, or are we gonna see a future where most things are very vertically integrated agent stacks. And just to kind of highlight for people maybe that don't aren't aren't as familiar with some of these dynamics happening right now, there's one side of things which maybe is lives in the world of it's sometimes more open, not always open, though, open source agent, harnesses, but things like, for example, Landgraf or Pydantic agents or Hermes agent or open claw, etcetera, etcetera. Name name your favorite one, framework or agent harness.
36:58
And within these frameworks or agent harnesses, you can interchange the model like the brain as as you mentioned. So that could be a closed brain model. It could be an open one. You could use multiple, etcetera, etcetera. But there there's a distinction there between that harness and framework layer and the the model layer.
37:21
Then then you can look at other stacks, like maybe you look more towards, in built agent features in Anthropic or OpenAI's platform, but also others, whether it's like IBM Watson agent thingy. I forget what it's called or, you know, the hyperscalers. AWS has that agent core thing, I think, like others. These are very opinionated approaches. Right?
37:49
To say, hey. We're gonna do all the different things of the stack. We're gonna do them in an opinionated way. We're vertically integrated. We maybe even have the compute.
37:57
We have the models. We have the agent harness. We have the the, interface. We have the console just like take everything in our vertical stack. Right?
38:07
So we we do see the both of those dynamics very clearly in the market right now. So what does what does that prompt in your mind?
38:15
Chris: Well, think we're gonna see all of them, because we're already seeing all of them, and I don't think that this is a winner or loser side. I don't think one falls away. I think that they fill different needs for different purposes, and in my own life, I use all of the above. It really depends on what my need is and what is the sensible architecture and strategy around what it is I'm building, and some of that is using closed frontier models that you would find at, whether it's the Geminis and the Claude's and the OpenAI models, and there's a place for them. And then others, I might be going for the JEMAs and the other smaller models where you can run inference on a device, for instance, or maybe an application that just doesn't need I mean, why spend the money on closed model APIs when you can do a lot of what they do at much, much, much reduced costs and host it yourself and get the economics on your side.
39:25
So it's an interesting question that you pose, because yes, all of them are gonna happen, but there's such an economic incentive for all the different versions depending on what it is that you need, And I think what we're starting to see now in recent days is really a drive. We saw the explosion of usage. We were talking about this a few months ago on the show, about just token maxing, and that has already It came and it went, because suddenly the cost of that caught up to people. So now what we're seeing is we're seeing people going, How do we get there at the most affordable cost for what it is that we need to accomplish? Which I think is a lot more sensible.
40:09
I'm glad we're past the token maxing moment into more of a sensible, Let's think about this architecturally and strategically going forward on what we need to do. So yes, all of the above with a sensible planning process guiding you.
40:24
Daniel: Yeah. And the you know, we when we have agents that could be operating autonomously for hours or days or months, obviously, then that makes a totally different impact to your your costs of operating that type of system than just a chat back and forth. Right? So that that's certainly a part of it. I I do actually think, you know, you mentioned that maybe both of these will exist, for different reasons.
40:53
In my mind, just to draw one of those reasons right now would be if you look at, you know, most businesses utilize some sort of drive workspace document ecosystem. Right? Mostly Microsoft or Google, although there's others. Right? So if you're primarily using Google Drive and Google Docs, then you're automatically gonna get one of these very opinionated AI integrations across those various apps in a way that is incredibly integrated into that, and that's Gemini.
41:28
But you would not be able to use that opinionated Gemini stack within Microsoft three sixty and documents and SharePoint, etcetera, because the one that is integrated in a very opinionated way there is, you know, Copilot, etcetera, across Microsoft applications. And so there's this suite of applications, and, obviously, they have their opinionated very tightly integrated way of doing those things. However, whether you're a Google shop or a Microsoft shop, if you're creating this, what I think of as your digital workforce of agents, right, you're not gonna have that digital workforce of agents be run by the little widget in your Google Doc Gemini widget. Right? And there's a whole variety of reasons that you wouldn't want to do that.
42:18
One one of them being, you know, just maybe cost and the the how much it would take to build that out. But I think to use a different I'm trying to think through all these examples right now because I love, I I do love teaching, and I'm I'm trying to think of better examples as as I teach more of this agentic stuff. But I think an example that I could use here is that may connect with some people is that most businesses, or many businesses at some point in the life of their business engage with a consulting firm. Right? Whether that's the big consulting firms like, Mackenzie or an Accenture or something like that or midsize or smaller or whatever.
43:02
Those firms have highly, highly opinionated takes on how to do things. They have their own systems. Right? If you engage with them, they're gonna do things in a very specific way to accomplish your goals. But most businesses do not say, we're going to staff our entire business and workforce through McKinsey.
43:24
One reason is no one would have that much money to to do that because it would be an ungodly amount of money. And two, like, what what what is your worth as a biz like, what is your what value are you providing other than the business entity? And that's where you can really build up the value of your business and your agility as a business over time, both from an IP perspective, but also your control, your flexibility, your ability to pivot and under economic pressures or under, marketplace pressures or pivots or that sort of thing. You wanna have control of that of that main part of your workforce. Right?
44:06
And I so I think that control and longevity element of being able to build agents within your digital workforce in a way that is definitely more vendor agnostic, less tied into the strict opinion of a vertically integrated agent stack, I think, does have advantages. Now will everybody be able to do that, or should everybody do that? I'm not necessarily saying that, but I I'm that's one example that I've been trying to think through as I'm thinking of what some of these dynamics are.
44:40
Chris: I think that's a great insight that you had, and that is having specific vertical stacks, because those are already developing. You named a couple of them. There are more out there, and they own a certain amount of the business. But I think the realities of business is that the ability to move across those verticals as well as into your own domain of your own the models and agents that you are managing on behalf of your business that are very specific to your business is a reality that everyone's facing at this point, and every aspect of business is going to have both the different types of models and the different types of agents having to do those interactions. So I think your point is really well founded that you can't think about just going and getting like, I'm gonna be an all in on Google, or I'm gonna be all in on Microsoft, or some stack like that, because I don't think in most cases it's gonna meet your full need.
45:42
I'm sure those companies will disagree with that, but the realities of life are too complex. So the key is to be I think this is the role over the next few years. There's a lot of people out there on LinkedIn calling themselves AI strategists and such, and if you fancy yourself with that title, then you've gotta figure out how to apply these different resources to the needs of the business given this very complex ecosystem. And so I think that is where things are going. I think it's messy.
46:18
Yeah.
46:19
Daniel: Well, I do think that, well, life is messy, so everything that we deal with in the future is gonna be messy, I'm sure. But there there is some there's definitely some patterns emerging that are helpful for people. I hope we've been able to highlight a couple of those today or even just clarify some terms. As we as we kind of get towards the end here, Chris, I think one question that might be interesting to just wrap us up here is how how should people evaluate their agentic AI efforts as they dig in and try to accomplish things with agents? Is that is that by, you know, how do they how do they measure measure success maybe in terms of a successful agentic adoption or that sort of thing?
47:11
Any thoughts?
47:12
Chris: I would start small with the expectation of rapid iteration, and just experiment your way into what's working. Be thoughtful about as we've talked about all these different options that are out there, be thoughtful about what makes sense. I see a lot of companies that are tied very tightly into specific vendors or vertical integrations, and all it takes is one of those vertical integrations to change something or to take something that may be a high value thing that your company does, and suddenly they provide that, and it can kind of crash a whole business. So there are times when looking at verticals might make sense, but maybe you should be thinking about how you can do it in a different, bound to a particular vendor so that your business survives as people are negotiating these things. So start small, experiment your way, and have a backup plan, maybe several backup plans as you're experimenting so that when you because you will run into problems, and you will run into limitations, and you'll run into models.
48:22
I run into models all the time on the frontier side that won't let me do things I wanna do because I'm trying to do novel things a lot of the time, and I hit guardrails all over the place. And so I'm doing that, and the other night I was sitting here at home working on a side project of mine, and that happened, and I'm like, Okay, well Fable won't help me, and I started downgrading. Have backup plans, figure out what will work for what you're trying to accomplish, and and see.
48:49
Daniel: Yeah. Great great way to close it out. Appreciate your insights, Chris, and it was a fun discussion today. I hope our everyone listening engages with us throughout the throughout the week until next episode. We are on YouTube now along with all the other socials, so go check us out.
49:08
You you may or may not want to know what we look like, but you can now if you want to. And so, yeah, we'll we'll talk to you again soon, Chris. Have a good one.
49:21
Narrator: Alright. That's our show for this week. If you haven't checked out our website, head to practicalai.fm and be sure to connect with us LinkedIn, X, or Blue Sky. You'll see us posting insights related to the latest AI developments, and we would love for you to join the conversation. Thanks to our partner, Prediction Guard, for providing operational support for the show.
49:42
Check them out at predictionguard.com. Also, thanks to Break Master Cylinder for the beats and to you for listening. That's all for now, but you'll hear from us again next week.