Ep. 7 What is Agentic AI, and Why Should We Care About It? | AI Insights and Innovation

[David Linthicum]

Introduction to Agentic AI

Would you like to create your own AI agents to do your bidding? You can with Agentic AI. Let’s talk about it. So welcome to the AI Insights and Innovation, your go-to podcast for the latest news, trends, and insights in artificial intelligence, including generative AI. I’m your host, David Linthicum, author, speaker, B-list geek, longtime AI systems architect, and analyst with ThoughtCube Research. Let’s get into the topic. So this has been trending recently, and I’ve done a few courses on this topic, Agentic AI, and I’ve been working with AI agents for years now. And it’s kind of cool to watch how this whole thing is evolving and kind of coming around, you know, out of the traditional LLM market, interest in generative AI, and now interest in creating and building AI agents, which is what Agentic AI is.

Defining Agentic AI

So what we’re going to do here is spend a few minutes defining what it is and how it’s different than traditional AI systems.

[00:01:07]

Applications of Agentic AI

Then we’re going to talk about the applications, and really, I’m going to end up telling you who it’s for, how enterprises can leverage this technology today, what purpose that it can provide, and where is the business benefit. So let’s get going.

Meaning of Agentic AI

So first, the term Agentic and what Agentic AI means. Agentic means with agency, and that means it’s able to operate autonomously and in an independent way. And we’ve been building AI systems using very similar approaches, architectural patterns like this for years and years and years. If you’ve heard of intelligence at the edge, and you’ve heard AI-enabled devices, all those sorts of things are kind of versions of AI agents and can be called Agentic AI. But there’s now some properties and some attributes that are coming to the concept that the thought leaders are promoting, and that seems to be bringing the industry around to defining what an Agentic AI agent is in the same way.

[00:02:05]

So when you look at what Agentic AI is, it’s built on the shoulders of things we’ve done in the past in building AI agents.

Comparison with Traditional AI Systems

We realized early on that the utilization of these big, huge deep learning systems and these big, huge machine learning systems were of some use to do very complex tasks. But if we’re looking for more lighter weight things, such as providing AI for a device like a camera or an autonomous vehicle, things like that, that those were normally too heavy weight. Also, they made decisions differently. And we needed something that was a bit more dynamic, a bit more iterative in the way it could work. It took also the ability to create many different agents that do things around their own personas. One can have a persona of a physician, the other persona of a radar technician, the other persona of a mechanic, and they’re able to collaborate together to carry out particular tasks by interacting or working together one to another.

[00:03:06]

In fact, you can build, and I have built in my career, systems that are made up of hundreds of AI agents that are working together to communicate, doing different tasks, different jobs, to build and deploy, to process systems, become part of business processes, orchestration processes, things like that. So it has a huge potential for the value that it can bring to the business.

Difference between Agentic AI and Traditional LLMs

So the primary difference between agentic AI and traditional large language model LLM processing, like chat GPT, exists in the operational capabilities and the frameworks that they use. Traditional LLMs, such as chat GPT, they generate responses based on the data they were trained on and are limited by their static nature. And of course, we can do things like augment the queries, the generation of the system using databases, RAG technology, things like that.

[00:04:00]

But pretty much they’re limited by what they were trained to do. So they don’t actively interact or adapt to external environments. And so they’re there to have a huge honking amount of knowledge. And we ask a question or we tell it to do something or we tell it to write a report or write a book or create a spreadsheet. And it’s able to carry those out based on the pre-programmed knowledge that is inside the LLM based on the training data that it was given. Now, there’s lots of different systems we can build around that to augment that capability to make it a bit more adaptable to external environments, such as integrating it with search engines and external databases, things like that. And that’s a possibility as well. And that’s an architectural option that we use when we build generative AI systems.

Agentic AI as an Evolution of LLMs

Agendic AI, on the other hand, it represents an advanced evolution of the LLM. So Agendic AI does process information like traditional LLMs, but in a sense, they’re small language models.

[00:05:01]

They’re more compact. They do things that are targeted at particular tasks. So they’re able to integrate tool calling capabilities that allow the systems to gather and use up-to-date information. And in doing this, they make more dynamic and responsive decisions and can respond to things and think through things and can consult with other agents and even reach out to communicate with LLMs and communicate with other databases and look at sensors to allow them to make a decision and also evaluate that decision to make sure it was the right decision that was made. So it does operate almost like a little mini human being that’s carrying out certain tasks and it’s able to reason, it’s able to refine its thinking, it’s able to look at external data points to make sure it’s making the right calls. So these agents can autonomously perform tasks and adapt to their environment by looking at these external tools or APIs, and they’re able to collect the necessary information in real time.

[00:06:02]

So the difference would be, and of course, there’s always a it depends situations, but that it’s something that normally carries out a certain narrow set of tasks via goals and pre-programmed decisions and objectives that you set within the agent.

Benefits of Agentic AI over LLMs

And using that, it’s able to leverage a knowledge model. Of course, it can be an LLM that exists within these systems that’s able to make decisions on your behalf or make decisions instead of a human being. And therefore, it’s a bit more useful than just a LLM, which is going to run on a huge server that’s allowing you to process these inference calls for something like this, since it’s designed to carry out specific tasks, you know, such as running a camera, such as autonomous vehicles, such as running a thermostat that’s on your wall. It can be designed to carry out very specific things and do so very well and learn as it goes.

[00:07:01]

In other words, it’s an artificial intelligence system, so it can build its knowledge model as it does as it experiences things in terms of its processing.

Planning and Decision-Making Capabilities

So also agentic AI systems can plan and decompose complex tasks into subtasks, providing accuracy and efficiency through continuous planning, reasoning, and a reflection process. In other words, it’s able to make a decision, but also look at that decision in terms of is it the right decision? Is this something that needs to be changed? And this allows them to work towards long-term goals, rather than just generating immediate responses like we do with traditional LLMs, like chat GPT.

Collaboration and Multi-Agent Systems

So they can also, as I mentioned earlier, work together, and you can leverage multi-agent systems to collaborate on complex tasks, overall robustness and adaptability. And that’s the cool part of this. And so we knew early on that, and I think businesses are going to find this out as well, that if we’re going to leverage AI effectively, there can’t always be these huge knowledge bases that we’re leveraging.

[00:08:09]

Tactical Orientation of Agentic AI

So they need to be tactically oriented at specific tasks or specific personas that they’re able to carry out. So in creating these agents, which in essence is an architectural pattern, we can make these agents, by the way, using the same tools that we use to build LLMs and other generative AI technologies. But in leveraging these agents, we can create these very cool, complex systems, which are many different agents carrying out many different tasks, operating independently one to another that are able to leverage each other to carry out and coordinate things like carry out a business process and how something is produced and manufactured, carrying out a process in terms of how an automobile runs, carrying out a process in terms of how a jet runs, and all the sort of core systems there, and get smarter and better at doing what they do and continue to evolve its thinking of not only the single agent, but the collection

[00:09:04]

of agents that are working together to carry out a specific application. Now that’s cool.

Components of Agentic AI

So what are the components? And like I said, I’m doing a bunch of courses on this now, so I have up to my eyeballs in understanding where this technology is going and also not only what it is and what it does, but how the industry is defining it now. And I think the industry is coming together on a set of attributes that relates to agentic AI. And so it’s the de facto way that agentic AI systems are considered. So the components are perception and sensing. They are systems that gather information from their environment, sensors, cameras, data streams, things like that, temperature sensors. I built one of these things for a pump jack, which is something that takes the oil out of the ground. And there was thousands of them. And obviously they were unmanned. And we had intelligent agents that worked within the pump jack that were looking at the temperature, look at the production rate, look at the oil flow, look for maintenance issues, things like that.

[00:10:05]

And in absorbing that information, they could make decisions based on the information they knew of that the flow rate was too high, that it needed to be slowed down. They knew the flow rate was nonexistent. There could be something wrong and that someone needed to be alerted. So it really is able to carry out a bunch of tasks, very much like if you had a human being that’s sitting there next to the device.

Information Processing Algorithms

We also have information processing algorithms, models used to process and analyze data, neural network pattern recognition. These things are able to make decisions, insight into decision making frameworks, rule based systems, machine learning frameworks, reinforcement learning frameworks. Again, they can function like an LLM. Most of them are going to be small language models are going to carry out specific tactical tasks, action execution mechanisms for executing actions, including robotic actuators, software commands, and then also learning and adaptability, the ability to learn from feedback and really kind of create their behavior or change their behavior based on things that they experience.

[00:11:06]

Learning and Adaptability

So, you know, if it’s a agent that’s running in a camera, it knows that if the humidity is too high, that the camera is going to be cloudy and how to clear that just by being told how to do that one time. And then next time the humidity is high and the camera lens is clouded up, it knows how to clear that. So the ability to adapt and the ability to overcome some of these obstacles is really the capability of these systems that is able to carry out things that normal static LLMs can’t. And so that’s why we’re considering the use of agentic AI agents.

Challenges of Agentic AI

But of course, there’s challenges. The technical challenges would be scalability. You have to consider the architecture and how they’re going to be able to scale. Now, normally agents are going to be assigned to do something that’s fairly narrow set of things.

[00:12:02]

In other words, they’re operating a single camera, they’re operating a single automobile. But when you’re building bigger systems, obviously, they’re going to have to have the scalability and the ability to get to their processing power and get to the memory consumption and the storage consumption that they need to scale those things up.

Technical Challenges

Data integration is also a limitation. You have to integrate these various systems. They have to communicate with these data sources. You have to set up middleware between the data source and the agentic AI agent, computational resources. They obviously, to optimize systems, make real time decision making. Things can be pretty computationally intense, even though I find you do not need GPUs to drive these things. Normally they can be CPU based and very low, low end CPUs. We talked here about Apple intelligence and basically an agent that’s running on your phone. And obviously that does not have a full set of GPUs in it like we would in some sort of a training system. And it’s perfectly capable of running in lower powered, lower intensive processing kind of environments.

[00:13:05]

Robustness and Reliability

And then robustness and reliability. These things are very complex and you have many of these different agents running on different systems, carrying out certain tasks. Obviously, it’s more links in the chain where things can go wrong. If an agent dies, it’s not carrying out a particular duty that it needs to carry out. The other agents aren’t going to be able to process or continue processing. That may be a blocking factor. So reliability can be a problem with these. And the more complex you build them and the more you build them with heterogeneity, in other words, different systems, different devices, different platforms, different operating systems, different cloud providers, all these sorts of things. And the more coupled you make them as they’re communicating and working together, then the more fragile those systems are going to be. A single agent goes down or a single agent fails some way. It’s going to stop the whole thing from processing. Of course, we’ve been dealing with distributed systems for a long time. This is nothing different. But you have to consider that when you’re sizing these things up, if it’s a business application that simply cannot fail.

[00:14:05]

If you’re building it using this architecture, there’s going to be a certain amount of fragility that’s going to be there. You’re going to have to put some operational operational capabilities in place, which works around that fragility, works around that complexity. And that’s going to be a little bit more costly than building an application using traditional techniques.

Applications of Agentic AI

So what are the applications for this technology? We’re seeing a lot of them today. Autonomous vehicles. We’ve talked about that already. They’re able to operate self-driving cars, drones, other autonomous transport systems. Again, the agent living inside the device. And I’ve operated them in drones. I haven’t seen what they do in self-driving cars, but there’s certainly AI capabilities in those cars now. And certainly they can be considered AI agents. Healthcare, applications for personalized medicine, robotic surgery, patient monitoring systems, which I think is going to be a big benefit.

[00:15:02]

So there’s no reason I can’t have an intelligent AI agent that’s operating on my wristwatch. Well, if I was wearing a wristwatch. That would be able to monitor my heart rate, my blood pressure and my O2 saturation, all these sorts of things, and relay that information back to other agents that perhaps existed within my doctor’s office. And they were able to coordinate certain issues and able to monitor my health and drive a better wellness process. And certainly we have those kinds of applications that are emerging today.

Use Cases in Different Industries

Many of them are going to be AI agent based. Industrial automation, smart manufacturing, predictive maintenance, supply chain optimization. Use that personally myself. And then personal assistants, which we’re going to see a lot of. They have powered assistants like Siri, the A word that I can’t say because it’s sitting right next to me, Google Assistant for various tasks and just really anything where it can benefit from the utilization of an intelligent agent that’s able to carry out a certain narrow set of duties.

[00:16:04]

Target Audience for Agentic AI

That’s really what it’s good at. It’s not good at something that’s going to be a complete replacement of a Binkhawk and LLM that we need for a particular business benefit. Things like that, doing things like major risk analytics where you need to train it with 100 petabytes of data. This isn’t that. Normally, it’s very fine grained, very unique tasks, very narrowly focused tasks where it’s able to do things on its own as an autonomous agent and also work with other autonomous agents to carry out some sort of a system application exercise. So how do you build these things?

Building Agentic AI Systems

Well, basically, like you build generative AI applications and AI applications in general. So you leverage machine learning and deep learning frameworks, TensorFlow by Google, PyTorch by Facebook, AI research, Kernes, API design, enable fast experimentation, scikit-learn, just any number of things that you can leverage out there.

[00:17:01]

Anything that can build normally a generative AI system or an AI system in general, machine learning based systems can build agentic AI intelligent agents.

Who is Agentic AI for?

So who is this for? It’s for product companies that are looking to build a product such as a smart thermostat, self-driving cars. A lot of the things we just talked about, drones, medical machines, personal digital assistants that is looking to use this technology and probably are using this technology today. I don’t think anybody out there who’s in that particular world is not leveraging intelligent agents, AI based agents to a certain degree. They may not be living up to what the industry is defining as the core properties of agentic AI, but they certainly have the technology in them.

Enterprise Use Cases

So very lightweight, very narrow use of AI applications that are needed by enterprises, by people who are building products that can use some of that technology.

[00:18:04]

It’s there as an option. Enterprises can use it for supply chain integration. The ability to build manufacturing systems are going to be much more durable, much more responsive and reactive via the intelligent agents coordinating together and working together and really kind of taking everything to the next level in terms of leveraging AI in application use cases that are going to be better adapted for what we need AI to do.

Tactical Focus of Agentic AI

And that’s going to be tactically focused at a certain very narrow problem set versus I think a lot of the enterprises out there are looking to replicate GPT within their firewall and they don’t need to do that. Normally, the use cases of AI, and we’re seeing this in the business, are going to be very tactically focused. They’re going to be narrow problems that we’re looking to solve. They’re not going to always be LLMs. They’re going to be small language models, tactical uses of AI.

Agentic AI as a Platform

And this is a platform. This is an architectural pattern we can use to make that happen.

[00:19:00]

So check it out.

Conclusion

So what I tried to do here is tell you what it is, what it does, what it means and who it’s for. And I think we’ve done that.

Future of Agentic AI

So check it out. I think it’s going to be one of these areas that’s going to get a lot of excitement and growth over the next couple of years. I think everybody, of course, has AI fever now. This is leveraging generative AI and new and I think more useful, more useful capabilities than everybody building an LLM, which I don’t think people can afford to do.

Tactical Use of AI in Business Applications

And so this is the tactical use of AI for any kind of a business application and the ability to use a very cool, very elegant and very useful, very adaptable architectural pattern leveraging AI.

Importance of Adaptable Architectural Patterns

And that’s what it’s all about. So anyway, thanks for dropping by.

Closing Remarks

I’ll see you guys next week. Take care.