Episode 103

AI Agents

Watch on YouTube

About this episode

In this episode of the Untangling Web3 podcast, we explore one of the most game-changing developments in modern Artificial Intelligence—the rise of AI agents.

No longer just simple chatbots or assistants, these autonomous systems can reason, take action, and interact with tools and data in real-time. We also break down the buzz around Model Context Protocol (MCP), the newly emerging standard that’s revolutionizing how AI agents connect to the world.

If you’re trying to understand where AI is headed next, this is the episode for you.

Key highlights

  • What Are AI Agents and Why Do They Matter?. AI agents are more than just advanced chatbots. They combine reasoning, memory, and tool usage to autonomously carry out complex tasks. From managing customer support workflows to making purchasing decisions on your behalf, agents represent the next evolution of AI—shifting from simple input-output systems to dynamic, goal-oriented intelligence. The hosts break down what sets agents apart and how they’re transforming industries ranging from e-commerce to healthcare.
  • How Agents Work and the Role of MCP (Model Context Protocol). At the heart of every AI agent is a reasoning-and-action loop—taking an input, breaking it down into steps, and executing tasks using connected tools. The game-changer here is MCP, a new open protocol that acts like a universal USB-C for AI. Developed by Anthropic and quickly adopted by OpenAI, MCP enables seamless communication between agents, models, and external systems. This interoperability is critical for scaling AI use across businesses and consumer applications alike.
  • The Future of AI Agents. Integration, Specialization, and Robotics: As foundational AI models stabilize, the real innovation lies in the tools and platforms built on top. Startups and big tech alike are racing to create flexible agent frameworks that can be tailored to specific jobs—from customer service bots to autonomous research assistants. The discussion also explores how agents might evolve into robotic systems, bringing AI from digital workspaces into the physical world. Combined with MCP and rapid improvements in model performance, we’re on the cusp of a world where AI doesn’t just suggest ideas—it acts on them.

AI agents are the next major leap in Artificial Intelligence, offering businesses and individuals scalable, intelligent systems that can think and act autonomously.

With Model Context Protocol setting the standard for integration, and global players like OpenAI and DeepSeek leading innovation, we’re entering a new era of digital autonomy.

Whether you’re building with AI, working alongside it, or simply trying to understand what’s next, AI agents are a trend you can’t afford to ignore.

This episode is sponsored by the VeChain foundation. Learn more about VeBetterDAO here: https://vebetterdao.org/

Transcript

Auto-generated from the episode audio.

Welcome back to the Untangling Web 3 podcast. In today's episode, we're diving into one of the most exciting frontiers in artificial intelligence, AI agents. Now, these aren't just smart tools or chat bots. They're autonomous digital workers that can plan, act, and learn on their own. From managing your inbox to powering self-driving cars, AI agents are quickly moving from research labs into our everyday lives. So, let's try and unpack it all together. Alec, I bet you're excited for this one. I am excited. I think um I'm a bit cautious but overall I am excited.

I think we've been talking about AI for quite a while now and it seems like there's always a new term or a new kind of hype coming out and I think this is something that we've seen a lot of over the last six months. We actually had Colin Fitzpatrick talking about and they're using this term like a year ago maybe so before it became the booming word that it is today. Yeah, I'm definitely excited to go into it maybe what it means, why it matters and you know maybe try and explain some of it as well.

Yeah, it's crazy how long ago it feels now that Colin was talking about this and now, you know, that's at the time wasn't really being talked about in most circles and now it's like the only buzzword. If you look at all the the kind of recent big tech company announcements and stuff talking about AI, uh they used to just be talking about AI and now they're saying agentic AI all the time. It's like the the newest newest buzzword. Um, so yeah, it's a good one to get into.

And I think, yeah, when Colin started talking about it, my interpretation was, oh, it's just a little avatar that can, you know, basically do a Google search. It can respond the way ChachiPT did at the time, but with a little face to it and looks a bit human, kind of like um some of the sci-fi programs that you see like Bladeunner. Um, but yeah, let's should we go into it like what we think it is and you know, what we why we think it matters? Yeah, it's a good place to start, I think, if it's just kind of defining it for ourselves because when I was thinking about it and kind of researching this episode, I had like an idea of what it was in my mind.

Uh the problem is it's kind of simple and complicated at the same time. So for me in the simplest form AI agents are basically just a wrapper around an LLM around a normal what we talk about as AI now gen AI uh language model but you wrap it with additional tools and the ability to essentially become task oriented. So instead of asking it to, you know, just summarize information or, you know, help you write an email or something, it's now goal based or task oriented. So it can try and achieve a goal for you, right?

That that's kind of the ultimate thing and how you get there. There's like tons of different ways to do it and you can equip it with lots of different tools, but at its core, the AI agent is is basically that it's trying to achieve a task, right? Okay. So even that I find that kind of like a bit vague in a sense, right? I find that kind of difficult to get my head around. And I think it's kind of like sim, you know, to go back two years ago to when we were defining web 3 and all this kind of stuff.

I kind of think of it in a similar light where we have this continuum similar to web one to web two where it's really hard to say where the line is where an agent is, an AI agent is, and where one is not. And I think part of it is there's this label for, you know, sexy investment VCs to be like, oh, look, we've got the new AI thing. Let's just explain it immediately what it means rather than be say some of the characteristics you said we've got a wrapper on a model that enables us to do XYZ and has these tools but yeah I think really roughly I agree with you in that we kind of think about you know AI models as quite static and agents are just a more dynamic form of that like they can take actions execute workflows trigger APIs adjust behavior and all this kind of stuff they're just a bit more dynamic they have more understanding of their envir environment.

When we kind of think of like the traditional open AI, it was just a context window of my prompt plus its previous static knowledge and it would give me something and I would do something with it. And now we're kind of thinking of that window of interaction of the world being much bigger. Like you say, having these tools, these access to live information, the ability to basically be more autonomous and have more agency. Yeah, I think it's a I agree. When I said it, it's kind of simplistic, but and it's also kind of a self-defeating definition because well, if my goal is to write an essay that's overdue for university, then you can definitely it can definitely achieve that goal for you.

But yeah, it's basically for me it's like anything that's a bit more complex than that. Anything that requires either putting getting additional inputs from the outside world or putting stuff out into the outside world and doing kind of more complex things, that's where that's where I think the line is drawn. Um, and typically, you know, when you look at like an LLM, um, that is, as you said, it's limited based on essentially the knowledge it's been trained on, right? And that's that's a big part of the the move to agentic AI is you want it to have more capabilities and knowledge context and that, right?

So, and also they're not always that adaptable on their own. If you just have an LLM, you're just interacting with it through, you know, Open AI through the the website and the chat interface, it's kind of it's not that adaptable to other types of tasks you might be able to think of. So yeah, that's that's a big distinction for me. And I think yeah, when you're kind of picturing this in your mind, you have like the foundational open AI model and it's built on top of a knowledge base and we can say that's kind of static, but it's really not because they're obviously plugging more data into it and it's learning based on your promise and all these kind of things.

But I think the kind of when we are to draw an architecture in my mind anyway of like an agentic AI the system architecture should change like it should be able to create its own links to new tools and pick say what APIs it wants access to based on not not even just requests from itself like requests from a user sorry it can basically derive what it needs to link to based on a really simple input and create more complex inputs from that. But again, this stuff is so vague.

It's just kind of a more dynamic version of the static LLMs that we've been become very used to over the last two years. Yeah, exactly. I think one of the best ways of describing it I think that I've seen all kind of ways of think again you kind of need a diagram for this topic unfortunately because you Yeah, it's the architecture that the agentic architecture is the key thing, right? It's how you build stuff around an LLM, just a simple LM. But the way I think about it is okay, people try and solve problems using LLMs all the time.

Like we said, you know, it's a simple example, finishing my essay on time, but people do more complex things and, you know, a developer like a classic developer might be able to hook up uh to the OpenAI API, right? They have an API that you can trigger events from in response to or you can use it as part of an existing workflow and often they get called AI workflows when you have an LLM involved. Um the big difference between all of that and what people mean by agentic AI I think is that whenever you have a workflow you have a human that's defining how that workflow works right you have a human deciding what's the logic that is going to use the LLM AI agents essentially just give that whole logic defining process to the LLM and say okay I'm going to give you a problem and you're going to decide the steps that you take to to solve it that's like the big distinction I think when you when you move to AI agents

and that makes sense from like an enterprise perspective, right? Imagine trying to sell this three years ago and being like someone's going to give an initial action to this agentic AI and then it's just going to go out onto the internet, start sending money, start, you know, opening APIs, retrieving information and all this kind of stuff. That's really scary. Like no company wants that kind of liability. So it makes sense the approach they went for where it's just kind of, you know, request an answer and then the human does something with it.

like I write me an essay, even write me a code based on, you know, these API keys I have, but I will be the one that's kind of liable if anything goes wrong because I'm the one that's going to copy paste that code and run that code and do something with it. Um, but now we're in the age where everyone's trying to be the new sexy toy on the block and they're taking a bit step further. Maybe they're trusting the models a bit more and saying, "Okay, actually, you're going to be given some directive.

don't know what it is and you're going to go out and start actioning things because now we feel a bit more comfortable that you're not going to mess around if we give you access to some keys and give you the ability to, you know, basically have payment rails and send transactions and stuff. You're not going to dump 1 million Bitcoin over here and all this kind of stuff. But yeah, it's quite exciting. I think the real edge here is the fact that it has a bit more agency.

It can do things itself. And the kind of the real cutting edge stuff is what can it do? How can it actually interact with the world? what tools, interpreters, whatever you want to call them, are basically creating that translation layer between the foundational model and everything that's going on in the world. Yeah, exactly. I think you've been ahead of the curve on this, Alec, because you've been talking about adding these payment rails to AI since like the start of the podcast. And I think, you know, in those early days, it kind of sounded like a bit of a pipe dream, but now it's very much a reality.

Like I think, you know, Coinbase, for example, have started implementing uh payment rails for AI. So, it's very much happening. Um and yeah, I think it's it's obvious why this is this has blown up as a as a topic in the last kind of six to 12 months. Um the impact it's going to have is obviously going to be big because if you can think the space of problems you can solve with a normal LLM, but then if you go and equip it with other tools, it's going to be able to solve a huge much much bigger problem space than you have already.

And there's like as as ever with AI, there's huge predictions like I've seen things like 5 billion in 2024 to an industry size of like more like 50 billion in 2030. I mean I I think that's probably an conservative estimate, right? I think this is going to be this is going to be huge. So like a nine times market growth in six years. That is bonkers. I think I kind of understand why. Like you could argue that the limiting factor with you know AI right now is that you have this human in the loop that has to check things is liable for things and a human can only work so fast they're not 247 obviously like AI is taken what a human can do and times 10 any productivity write me an essay based on these things do the code to do this but in a world where it's AI interacting with AI how we think of like the economy commerce communication it is going to just we I don't think we'll be able to

comprehend in 10 to 20 years time. You know, you see these like crazy stories like three, four years ago where it's like, oh, they put two AI models in a room and they made their own language and within a minute, no, no one could understand what they were saying, but they were obviously transacting something useful. It's going to be that kind of thing where we won't understand the global economy in my mind in like 10 to 20 years because the economy is going to be based around AI interacting with AI because that's the most efficient and effective way.

And that's why there's these predictions on such huge economic growth. Uh yeah, I think we'll see what happens, I guess. Yeah, I mean I've seen in the last few weeks a bunch of I think it's Shopify, might be wrong, but there's basically a few different um kind of platforms around online commerce who are shifting to we expect to be, you know, receiving money from you via an AI. Your shopping will be done via an AI in the future, right? And they're actually gearing up for that future.

So, you know, instead of having all that decision anxiety when you want to buy a new jacket or something, it's going to know your preferences and it's going to go out and buy things for you. And I I still can't quite get like that little leap in my head of like letting someone else do my shopping for. I don't like getting Christmas presents that are clothes from other people, right? I can't imagine how that's going to work. But that's the kind of thing that these big companies already moving towards and expecting, aren't they?

You're easy, mate. Your model is just a medium black t-shirt like the same thing every week. Yeah, it is bonkers. I give the easiest AI agent jobs ever, so it's fine. Exactly. I can understand it for things where it's like food shopping, easy to do. Buy milk. You know, that was the classic IoT example. It's going to be hooked up to your crypto payment rails or when your milk's running low, you it's going to automatically buy your milk. That's crazy. But the idea of this being like the interaction layer for buying clothes and I think it's kind of scary.

Like I think it takes a lot of the the power and decision- making away from humans and we can't understand things and when things go wrong, how are you going to explain this stuff? But yeah, we've got ahead of ourselves. We were just meant to try and define what it is and maybe we can go into some of the the background history and context for all this stuff as well. Yeah. And I think um you know, when you think about the history of AI agents, I would I I would say the AI agents is just the the last piece of the history of AI, right?

It's the most recent thing that's been happening in AI. because of how we defined it and I think this kind of slightly vague definition that you have to apply in this case um you could argue that there's been a bunch of milestones in the last kind of 50 60 years that would be they would fall kind of under this definition of agentic AI it's like computers that could automatically make decisions uh based on you know a bunch of external tools and context and things so you know when we when we've covered AI in the past we've covered that full that full history so I don't think we need to do all of that but if you look at something like uh you the IBM computer that beat Gary Kasparov in the chess game.

You know, that's that's an example of solving a kind of non-extbased problem. It's a different problem that was able to do with a bunch of extra different context. Uh similarly with IBM's Watson, I don't know if have you ever seen it the the episode of Jeopardy where IBM Watson competes. It's so good. It's it's so it's so impressive. I think like it we don't you say you don't we don't need to go into too much detail but I think the earlier agents were like again it kind of gets to this web one web two thing were really rigid they followed rules they didn't adapt very well like good at chess is but terrible at real life right so when the rules are very defined the models very defined easy but not so impressive when we talk about like really complex abstract maybe unquantifiable models sometimes for for the real world even Eliza back in like the early days 1960 whatever it seem really impressive but it was just mimicking language like it there was no

real kind of understanding or understanding of nuance there and it's hard to know what understanding is right this is problem with AI it kind of gets so philosophical a lot of the time but yeah Watson's a good example Alph Go is a good example of when it started to really you know understand nuance understand risk which is really quite complicated when you think about it understand strategy and all these kind of things but then like we said when we started talking about you know open AI and chatbt in the early 2020s like their big leap was this conversational ability like the ability to have human prompts and understand nuance in language and patterns in language and be able to like create useful outputs using that and also like task chaining was another big thing and like a a thing we've seen in the last year that kind of relates this agentic AI kind of development and trend is the integration with external tools as well being able to take what's been done before have the

foundational model that understands language understands patterns and plug that into a tool that say get in lots of data to create useful outputs. Yeah, I I agree. I think you can kind of just use the late, you know, uh, sorry, the early 2020s as the starting point for AI agents like that we know of today, you know, in many ways because they're so deeply embedded uh, into the LLM model, right? It opened up the space of of language that it could process that you you can define tasks now.

Um, so it's everything all the AI agents are have been based off that partly because it's being so usable as well, right? If you're going to make uh the kind of holy grail of general purpose agents that can go and break down any task you give it and if you give it the right tools, it can go and you know achieve that goal or at least get you most of the way there. Then all of that is being built on the most successful version commercially of of language processing which is LLMs themselves.

Um, so yeah, that that that's I would just, you know, you may you can go back and research the whole history, but just call it like, you know, 2020 2021. That's pretty much the starting point of what people talk about now as AI agents. Um, and yeah, you mentioned I think we'll get on to the specifics of how they work in a minute, but yeah, that chain of thought or chain of reasoning uh model for how LLMs work because that is yeah, that's pretty recent. That's only in the last kind of I guess year to two years that that's been around.

But for anyone who doesn't know what that is, if you now go and use like chat GPT and you can see the it will now t typically show you the reasoning process it's going through. If you use deep research for example, it'll it'll break down the steps it's going through for that researchbased task. So that's kind of the the an analysis of your ask of your problem like go and do some research. It breaks down okay well I need to look at this. I need to look at these websites.

I need to synthesize this type of information and put it out in that format. So it's already you can already see it in a simple narrow problem set but uh yeah this is now being widened up to kind of basically anything that you can give it a tool to go and do. Yeah, I find that that process of seeing like exactly what stage that's so satisfying. But again, a lot of that comes down to context window like the the amount that these models can keep track of at any one time.

And early on that context window is quite small. So it was more like um I saw it being described as like quite synchronous task. So singleshot question answer like you know why is the sky blue? Ba bum it answers it. But then if you ask like okay well why is know why is it sometimes not blue? it would just start again because it wouldn't be able to keep tabs on everything it already knows and they the question before because then the context window becomes too big for it.

But with as the context window grows, we're starting to see say, and this relates to AI agents as well, these kind of asynchronous tasks, the agents kind of running in the background, monitoring systems reacting to change, kind of like a digital co-worker. Whereas before, like I keep coming back to this, but I think it's really important. We think of AI as these containerized windows of interaction. You ask a question, it gives you an answer, you go do something. But more and more we're seeing it being an open world whereas like a digital assistant you give it an agenda or even just an aim and it will go away for a day, an hour, whatever it is and just do lots of things in in the background that you wouldn't be able to have thought of in advance yourself.

But as it's kind of, you know, finding bottlenecks, finding problems with its say initial kind of idea of how to do this, it starts to create these decision trees and it goes off at a tangent and maybe you know paperclip factory ends up destroying the human race. Yeah, it's it's a good way of thinking about it. Actually, I hadn't really thought about it in the asynchronous way, but yeah, it makes sense. Like, you're just kind of you're duplicating what you can do in a given window of time.

So, the ultimate problem you're solving with agents is just time, like things you don't have time to do or things you want to spend that time doing other things with. Um, so yeah, I I like that. Should we get into how they work? Because I think at a high level, they they're actually fairly straightforward to understand if you don't go too deep into the details. So, at a really shallow surface level, so easy to understand. There's like a PhD in like deep learning right now being like no they're freaking not.

Go on mate you go. This is very it's nice. So if you they the experts have kind of condense things down for us normally into what they call the react framework which doesn't help for developers because react framework means to something totally different in front end development. We'll just ignore that. Um and react standing for reasoning and then action. So that's the two kind of the two main um I guess domains in which these these AI agents work. So the reasoning part means they have a almost like a perception.

You can give them an input, a prompt much like you do for uh you know chat GPT. It's the classic way of using NLM. You can create an instruction or give it a problem to solve. The reasoning is that kind of chain of thought we've just talked about. So it can reason not just you know uh how to uh write an essay in this case. It can say okay you give me a complex problem and that may be very complex this problem. Then it reasons okay what are the steps I need to take and it will write down and think you know in its own memory essentially what steps do I need to take so that's the first part is reason and then the next part is action which kind of speaks for itself but that's where it goes away and takes those steps implements things and then uh hopefully it gives you the right answer at the end or some answer so that's at the high level that's the key of of of how

AI agents work right today's episode is sponsored by the Vchain Foundation now we often talk about how emerging technologies like blockchain have the potential to revolutionize sustainability and Vchain is at the forefront of this green movement with initiatives like V better DAO. Let's take a closer look. Imagine tracking small actions like recycling at low cost making them valuable assets. These verified records build a massive collective impact benefiting people through their own actions. Sound a bit complex, but Vchain made it fun. Explore our apps at vb better.com.

Vchain's goal is to power a new era of sustainable solutions. So, if you've got an app idea or just want to earn rewards for small sustainable actions every day, go check out vb better.com to find out more. And now back to the show. And I saw um like another way of describing this is like a a pipeline of like let's say the steps similar to what you were saying where the first step was around perception. So it's taking in data from sensors maybe from APIs like understanding what's going on online could be Twitter sentiment all this kind of stuff.

Then the second stage is what you mentioned there is cognition using models to reason plan understand really what's going on like all this you know the many huge amounts of data that it's taken in actually turn that into something useful maybe a planer strategy like you were saying there based on a request and then yeah the final step is actioning that performing tasks sending an email booking a meeting writing some code developing something that kind of thing but yeah it sounds really simple when we're saying like that that's kind of like like how humans act as well like how humans interact.

So I guess we explain this in a simple way because it kind of seems more intuitive to humans and I imagine again there's probably some expert listening being like that's really not how it works. Yeah, exactly. I think um for our small brains it's kind of smooth brains even that's the easiest that's the easiest way to think about it. Um and again that reasoning part is yeah super complex like how you actually get it to reason is you know beyond the scope of my understanding right now.

I don't know any anywhere near enough detail, but at least you can see it working, right? If you use an LLM, the action part is kind of easier to grapple with because that's actually more about just connecting it with the real world with, as you said, APIs, with data, with real time data, for example. So, that's the difference between, you know, in the past, like a year ago, we were saying, well, it's a big limitation of LLMs that they don't have up-to-date information or they're not trained on up-to-ate information, so they don't have access to it, right?

So you've seen we can kind of get into retrieval augmented generation or rag as well when you can just kind of have an an LLM look at a certain data set that you've got like when we talk about a private company uh building a kind of second brain for its uh based on its own internal documents that's like a very small subset um example of what an AI agent is right you're giving it a certain set of inputs but yeah it could be literally anything it could be code computation calculation uh B or other LLMs, right?

It can be other agents, which is the kind of future that you're hinting at, right? I think that we're going to get to is actually what we're going to connect it to most likely as well as just Google is a bunch of its friends, it's other AI agents in the in the ether that are yeah, talking to it and helping it solve certain parts of problems as well. Yeah. I think like one of the so the the kind of the early advancement I kind of think of this as two parallel streams of advancement.

one is like the foundational model, the brain that sits in in the middle, the one that comes up with the strategy, the cognition, whatever, all that kind of thing. And we're seeing like there are a few big players, maybe six or seven. We did a whole episode on it last week, couple of weeks ago, on the advancements and how these things have, you know, they're more performant, higher context windows, all this kind of stuff. And there's kind of like two branches and on and how companies are approaching that.

One is just throwing a huge amount of compute to the problem. That's kind of like the X Grock style which we didn't even talk about in the last model. They're just throwing so much compute at the problem and it seems that they can just brute force it and create really good understanding responses whatever with a lot of compute. And then you've kind of got the deepseek model which is think basically kind of breaking down questions and answers and patterns into a more like smart way, efficient way.

And like I said, deepseeek is extremely performant based on the amount of compute. Like you can run the DeepSec model on a MacBook, you know, which is bonkers. Like really bonkers to think that's possible. And like these companies, you know, it's getting better and better. And it seems like the the cost of compute, the cost of cognition is coming down and down, which is like kind of commoditized in a lot of ways. It's quite exciting. And then the other thing is the tools built around that big brain.

And you got the inputs and the outputs. What am I seeing? What is this information that's being fed into my brain? and what can I actually do with that information when I come up with the strategy? And I think that's all to play for right now. Like all the rappers you're talking about, the specializations that we're talking about, that is kind of where I see there's a lot more maybe opportunity for startups. They're like specific use cases, specific angles, specific assistance, all this kind of thing.

But yeah, they're the two big streams I kind of see being developed right now. Yeah, exactly. And that's that's a good point, right? because there are lots of different types of AI agents, but I'd say another way to dice them up in general is that you have, I guess, specific or special purpose AI agents that are people again people like you see it like uh marketers for example, they are there's loads of marketers currently building AI agents for marketing, right? That can go and send cold emails and write copy because actually that's kind of stuff that the LLM are pretty good at already and it's a case of just connecting them via Zapia or something like that, right?

So that's kind of simple. And then the other branch is people are pushing for the general purpose AI agent that can kind of understand and do anything that we'll we'll talk about that a bit later when we kind of survey the landscape. But there's one called Manace at the minute that's come out recently that everyone's kind of going a bit crazy about because it's it's it's meant to be general purpose and it can handle a lot of things and it's doing some some interesting things behind the scenes for how it does that.

But you know again to to like in some of the categories you have of just like simple agents this shows the difference between uh what I think LLM based agents are in 2025 versus kind of what they've been. So you have like these what they call simple reflex agents which can just do simple things respond to inputs and they're not actually that clever like you know like a special thermostat or something like if the temperature uh drops below this uh value then then you know turn that temperature up in the house like that kind of thing.

It's very simple but that's an example of almost like an AI workflow more to me right where it's that there's logic that's been programmed by a human. I've decided that's the behavior I want. And then you have everything all the way up to these learning based agents that can learn and improve over time where you have the agent itself. The LLM in the middle of the agent is defining what the logic is, the control logic. So it doesn't it's not the control logic isn't if this happens do this.

The control logic is like a dynamic thing that's being updated. It's the LLM deciding okay I had these instructions last time this kind of didn't work that well so I'll change the instructions try again. That kind of thing. They're very iterative when they work as well. I think it's quite un important to understand the different types of like these agents based on your use case. So like I mean we haven't mentioned all of them but say a HR agent might be a goal-based hire like that's another type of agent a goal based agent like get me three CVs by the end of the month are all super appropriate and you know make sense for this job and that kind of thing.

You might get a utility based agent which is another example which is like pick the best candidate from those CVs that you know the first agent's been given and then you might have a learning enable one which is like improve this process over time or help say um attach to an individual an employee and improve their learning over time and like when we're kind of thinking about this you know you have the the kind of philosophy there's a general agent that can do all those things perfectly but more and more we're seeing these specialized agents and kind of thinking about this multi- aent world in which you'll have Okay, those three specialized agents I just talked about.

So the goal based agent that finds people, the utility based um the utility based agent that picks the candidate and the learning based agent and they're all specialized in their very own like specific ways and they interact with each other and kind of feed off one another because they know that say they can take over the next bit and that's their specialty. It's kind of like what we think experience right now with like globalization, capitalism, all this kind of thing. No one person does it all.

we have very unique specialisms say geographically or kind of whatever sense and we rely on other people to be very good at the thing they do. Yeah, it's it's a really good point. I think um I also want to touch on the utility based agents thing as well because that's it's really interesting because how you train like a traditional neural network or an LLM itself is you know not to go into all the detail but it does use something called like a cost function or utility function where every time it's improving the model every time you're doing this training and trying to get an output based on a set of known inputs and outputs then it it'll give you an answer and it essentially rates that answer and all these kind of weights in the neural network they updated based on how expensive that was, how costly that answer was, how close it was to the correct answer and that's how it gets updated and the very same thing is actually happening now with AI

agents in the space of like how do you solve a problem. So when you give it a problem I said it it'll do the reasoning it'll break things down into steps but obviously there's like so many different combinations of steps that you could take for any given problem. So it will build its own kind of network of thousands and thousands of paths that it could take to go and solve this problem and then over time as it solves more and more problems it will decide what's the best path for solving that problem.

What's the best set of instructions I should follow as the LLM as the AI agent to go and do that. So it's kind of that's the it's the kind of how they're learning to be better at solving tasks as we go. It's very meta, right? Because we train the agents this way and now they're updating their own processes like that which is is a bit spooky. And that was one of the interesting things, right, when open AI and chatbt initially came out is like people were giving negative feedback.

So they were saying like, oh, it's actually our models are actually kind of um performing worse over time because people are saying, you know, giving down thumbs to good answer and up thumbs to bad answers and like feeding in wrong data sets and all this kind of stuff. And like you say, more and more we're not trusting humans to necess necessarily do the validation. We're maybe trusting models to validate it. And there's easy ones to validate like say a deepseat model might be like the the function that I'm trying to optimize for is cost or cost to compute or energy.

That's very easy to do. Like basically, you know, we can work out how much computations involved fairly easy for a mold to do that. But in terms of like quality of the answer of like how good is an essay, like that's really quite difficult. That's very subjective. Like how do they monitor those things and measure those things? And yeah, maybe we maybe we'll get into that, but that's quite a complex one. Yeah, exactly. Um should we just survey a couple of the major ones that people will have heard of I think because that again will make it a little bit more concrete.

So, I guess the one that beyond Colin and and kind of the some more niche kind of companies, I think people started talking about AI agents a lot more when OpenAI brought out operator, right? And I think it's still in like a preview, a research preview essentially. I don't know if you can access it fully, but essentially, yeah, this is their again very it's almost like it's not that it's not like a a full bells and whistles version of an AI agent. It's very they're very good at productizing.

They make rolling out slowly and carefully, but it's essentially what we've been describing. It's chat GPT powered with additional capabilities, right? So, you can build virtual assistants off them. You can automate kind of business tasks based on a set of a set of rules like that. The open AI operator model an agent is kind of one of the very first steps. I don't know if uh yeah, if you have any thoughts on that one. Um, I've seen like I think there's so many companies that are trying to be the new, you know, kind of disruptor and have their own AI agent right now.

Like the funding involved is crazy. I think it's like based on what you just said it's important to highlight that all the major companies like the anthropics with Claude and Gemini and all these kind of things they even in Microsoft they all have their own play for enterprise level AI agents to basically report replace either researchers developers admin administrative staff operational staff whatever it is they're all making big plays to hit big institutions and enterprises on that but then you have a lot of smaller companies like I said they're basically just taking the foundational model from these big companies which obviously they should do maybe deepsey because it's so cheap and private and all this kind of stuff and they're building this rapper in these tools that is effectively targeted towards creating a new AI agent.

We've got crew AI, autogen, they're all doing very different things. First, customer support is quite an obvious one. Like software development is another big obvious one where these things can have huge amounts of impact because 247 super efficient incredibly patient like it makes sense to have these customer agent bots that you know even if it's just to get rid of or you know deal with 99% of the really simple queries and you push the 1% to humans that then can answer those questions more effectively with the time they have.

Yeah. And there are like a bunch of essentially agent platforms now, right? which is I think that's a marker of where we are right now in 2025 as to where we were when Colin mentioned it because when Colin talked about it in that episode we also talked about there's like a there was this developer who's quite famous on on X on Twitter for like he was building agents on his own basically he was like building agents that could spin up websites based on uh talking to Siri like he'd basically turned Siri into an agent which is really cool and like everyone was mind blown now that kind of stuff is what is being productized by lots of these companies you mentioned like the crew ais autogen for example langchain they're trying to build systems to make those AIs and as you know we mentioned earlier Manis I think that's a really interesting one because you know everyone I say is is getting very excited about that being a general purpose uh AI agent builder and

that is uh the way it's working and the way they've achieved that is interesting like because it's bas basically exactly what you were saying just before Alec right it's they have like a bunch of different agents working in concert when you create one of these like kind of bigger agents, right? These general purpose agents. So they have like one agent that does the planning that takes the task and does the reasoning and then it has like four or five different agents underneath. One is like for managing memory, so what parts of a conversational context it remembers, one is for searching for information in real time, one is for, you know, various other things, API connections, right?

So it's it's like it's completely compartmentalized it and everyone's getting quite excited about that. Can I say I think that is such a huge play. Like we kind of talked about very specific use cases and you can like get really excited about this stuff like healthcare agents, content creation agents, customer service, you know, coders, all this kind of stuff. I think one of the biggest plays is having this kind of like centralization points because these models are going to change con like constantly, right? The models are going to be constantly improving and then if you think of that as the brain and these companies just developing the tools around the b brain as the models improve they'll the tools that are built around the brain they might need to replace the whole brain that might be complicated for the company the company could go under whatever that's a lot of dependency on these foundational models if something goes wrong for example um so I kind of see like the idea of having a degree of

flexibility interoperability like this translation layer like you kind of mentioned an example there another really good example is AWS s like they have this solution called bedrock where you can effectively define the solution architecture say the data inputs the knowledge base say the queries what you want and then you just plug and play and select whatever foundational model is hot and sexy right now and I think those kind of like applications are you look at AWS one of the most you know successful systems in the world right now for software development and solution architecture like we're seeing a similar things where there's going to be such a a plethora of foundational models and tools being developed constantly One of the big money plays is to be the aggregator of all those things that a user can come in with and be like, I'm going to choose the model that I like.

I'm going to choose the prompt I like, the knowledge base I like, and if I want to change it, you've got the solution where I can just plug and play and change to the new hot and sexy model. So, I think yeah, that's a really exciting play. I think a good example is AWS Bedrock. If you haven't checked it out and you are building these things, have a look because it's so intuitive to use it. Yeah, that's like I think that's such a great point. like the holy grail is going to be not just it being general purpose in what it can do but also having like maximum flexibility of what models you can use and what systems you can use having kind of universal integration.

Um speaking of which I think that's actually a really good segue on to kind of the next thing we should definitely cover on AI agents which is another acronym MCP model context protocol. So that's again even more recently than AI agents that's become like the newest of the new buzzwords right in the last kind of few months and to put it simply it's it's a protocol and framework that helps part of this AI agent workflow that we've been discussing right it helps with the action part so when you're connecting your agent to the outside world at the minute it's like it's like 200 or what five or six there's like USBA USB you've got all these other weird and wonderful cables.

Uh, and you always got the wrong one in your bag. Essentially, the way people describe MCP is it's like the USBC for AI agents, right? It's going to be a a universal way of connecting agents to external sources that everyone can use and everyone's just loving it, right? Everyone's going crazy for this. Yeah. I think like an an example that I've seen talked about in in the kind of a similar light is that AI systems today are like pre smartphone gadgets where every integration is bespoke and this MCP is like Bluetooth that kind of has this common architecture this common communication protocol that allows all these different agents being developed completely isolated completely separately for very specific use cases sometimes to plug and play access tools databases services all communicate in the same way I guess that the key thing is that these standards get adopted and all the foundational models and all the tools and AI agents are built to those specs.

But yeah, when we talk about this world in which I want to choose an agent and then plug into a database and define my specific use case and if that agent's no longer the hot topic, I can just plug and play something else. You need things like these. You need these communication standards effectively that everything is kind of, you know, judged by, I guess. Yeah, it's it's it's a really interesting one because when you talk about standards like yeah most most successful technologies need standards like the web two needed standards for how HTTP works right all of that is publicly standard all the encryption we use online has to be standard and publicly known and so everyone can use it in a common way and I think it's it's probably going to be I think personally MCP will be one of the fastest adopted standards in history just because the whole point of using an agent is that you're not meant to do.

You're not meant to put the human back in the loop, right, while the task is being solved. So, you don't want an AI agent to have to go and do an API integration. Like, that's that's that that's something that okay, maybe it's going to be possible, but it's like it's adding complexity where it doesn't need to be. And you want to be able to say here to this agent, go and solve this problem in the knowledge that it's going to be able to communicate with whatever it needs to, which is exactly what MCP is doing.

is basically saying that whole headache of all the, you know, weird and wonderful mismatches between different systems, this should be taken care of by MCP. You can just fire at it the prompts you need and hopefully get the right answers. So, yeah, it's it's it's going to be I mean, obviously it's going to be adopted quickly, I think, because it came out from Anthropic and Open AAI, one of the well, their biggest competitor arguably immediately added it to their, you know, website and their standards and said, "Yep, we're happy to use this." they know the power of interoperability and having the global standard that everyone uses.

Well, it's an interesting one because we talked about like the USBC standard and all this kind of stuff and the universal charging ports and Apple resisted that for so long because they kind of had this monopoly and my instant reaction was surely these companies will resist that. But I think the point is that you know foundational models are no longer the hot topic. The topic is the tools and the AI agents that are now being built. And for any of these companies to be competitive, they have to be as plug-in and play as ready to integrate with all these tools as possible.

So whatever the standard is, they're all jumping to be, you know, the leading model that is plugged into these standards because it's all about the tools now. It's kind of funny to be like foundational models don't matter anymore. It's about the tools the foundational models are built on that really is exciting people, which is crazy to think of this only been around for two years as well. Well, and when you've got such a crowded marketplace for the models themselves, there's enough ways that people are already using to try and differentiate between each other and it's it's a fierce enough battle between them.

You don't want to go and put up a wall garden around yours because everyone's going to be like, well, no, why would I bother? Right? Because yeah, the tools are going to make this model much more usable. Even if it's not quite as good on the benchmarks, it's going to help me do what I need to do faster and easier 99% of the time. And I think a big part of that is as well is that that what we've kind of seen is that the foundational models the upfront cost for a lot of these foundational models is so huge like it takes so much money so much time so much energy to really build and train these initial foundational models and then once it's done a year later the price for someone to just copy paste that foundational model or do it in a slightly more efficient way is like you know 20 times cheaper.

So maybe the companies start to say actually the foundational models are good enough now that average IQ is like 110 or something like that. So higher than the the human average. So they're functional enough to replace a lot of say human productivity or human labor, human intellect, whatever. Now the real money maker is in the specific tools that we're building on top of. And maybe that's where they're seeing the impact like you saw what happened with Deepseek when Deepseek came out with his free model. Alibaba did the same thing.

I think the big western, you know, US AI giants right now are being like, actually, it's gonna be really tricky for us to monetize these foundational models when the Chinese companies are just bringing out something as effective for absolutely free open source that anyone can run. So maybe the real money maker is these AI agents that can replace individuals, but we just got to build the tools to make sure that it actually works. Kind of scary, kind of exciting. Unsure what to feel about it. Yeah, exactly.

It's like it's almost feels like everyone's going to have to go for a platform play again. like you know we have software as a service for all these big companies that's how they got it so far. You want to become the platform for building agents right and not have to compete on what model they're using underneath like give the users the choice but you just need to make it as easy as possible to create really effective agents. So yeah it's going to be interesting to see how it plays out.

I mean the impact is obviously going to be huge like we've talked about the impact of AI at length and I think anything we say for AI goes again here but like tenfold right because now we're talking about it doing all these crazy things like doing your shopping uh it's almost like it's like finally the AI is on the other side of the coin for the consumer as well because for years we've been advertised to by AI or recommener systems and you know you have Amazon saying would you like to buy this again on a recurring basis and now we're going to be putting an agent in front of that and saying, "Yeah, you you can handle that.

You can say no to Amazon when you like, hopefully." But yeah, it's what are you are you worried about the future with agents or are you excited about it? Um, it's hard. I think like because I'm in the space, I feel excited. I think if I was someone that was, I don't know, one of these jobs that's going to be disrupted by agents, say like marketers or something like that or social stuff like very, you know, kind of I guess the sectors that are very information in, information out.

say if something goes wrong it's not the end of the world like it's not like a nuclear red button style I would be slightly worried I think that we're seeing AI agents very targeted towards support centers call centers um you know maybe like all yeah any kind of supportive like kind of in informative things like that I'd be very worried but because we're in the space I find it quite exciting because like the things that they can do and the kind of how pervasive they are in life is becoming more and more extreme like we're even seeing Alexa plus is now the new thing like imagine Alexa that's one of like the most hotly used AI systems in the world.

There are hundreds of millions of users of Alexa, but it's so dumb. Alexa Plus is now being rolled out. Imagine like, well, Claude is now going to be built into Alexa. Imagine the intuitive responses that you get from your prompt right now being built into Alexa. Like, it's going to be absolutely gamechanging. I think like the productivity and the ease of day-to-day life that's going to come from this is going to be so incredible that we can't even imagine what the next 10 years look like.

I think there is a question around governmental response to this and what's going to happen in terms of employment unemployment. I think we're going to see unemployment shoot through the roof because you know the classic kind of response from governments when you know the economy is not doing too well is to lower interest rates so people can invest more and then get more jobs but when the reaction to lower interest rates and people having more money and more debt and more loans and all this kind of stuff is then to buy more CPUs that doesn't correspond to more employment.

So what's going to happen there? So I think yeah excited and nervous I guess and daunted. Yeah I think um I'm I'm kind of in a similar headsp space with it. I think there are there are some really interest I think there are huge huge benefits to AI agents that are actually narrow focused. So I know I've been banging on about general purpose AI agents and that could be really really useful for the people who can use them well but having a narrowed scope back again of an AI agent.

So narrowing what it can do for safety reasons means you can siphon it off for specific tasks that currently are, you know, take too much manpower to do essentially, right? And and there there are plenty of things that aren't quite there's kind of a sweet spot between how critical the task is and how how much of that task you need to do like what the volume of requests for that task is. So there's obviously going to be applications in, you know, things like healthare for example where if you can pick certain non-critical task maybe paperwork based type stuff.

Yeah. Um but again it's about how how much do you limit the capabilities the sign off type uh limits of an AI to do things like you can limit some menial tasks but you also don't want it like you know doing all the prescriptions for example right because that's where you don't want any errors to to happen so yeah I'm kind of yeah it's it's going to be it's going to affect basically most industries I think you could it's probably good to have a hight touch kind of high nuance job where you where people still want people involved uh because definitely like high-risisk jobs as well, right?

Like where you need you need a human to to take on some of the risk, right? Because that's a whole other conversation is who's going to bear the responsibility for when things go wrong. Um like with self-driving cars, you know, that kind of stuff as we always say. But yeah, it's it's it's going to be interesting either way. Yeah, I saw um I saw some econ economists say that like most uh politics right now is left and right, labor versus capitalism, that kind of thing. said more and more it's going to go in when like you know AI interacting with AI and all the money is controlled by AI it's going to become humanist versus AI focused and like that's going to be the new political domain which is really interesting I think overall like it's going to it's obviously going to have huge ramifications positive ramifications for like productivity for food creation for health care for education all these things that are really maybe like basically kind of quite inflationary right now um I

think that you know We're going to become we're going to understand and have be accountable for less and less of our lives for less and less of like our systems like complex systems especially I think the real gamecher and it's kind of obvious is like when we start to tie this in with robotics and you know when you can buy a robot for five grand and it costs a pound a day of electricity to run and you're also plugging in these intelligent AI systems on top.

What does that mean for us as a civilization? Like I'm reading Isaac Asimov's uh robot series and I've already read Foundation and it's kind of like quite dystopian. Oh, it's dystopian and utopian. You can find like elements of both. I think it's exciting. I think I am overall positive about what's going to happen. I just think we can't possibly imagine what 20 years looks like from now, which is a crazy. Yeah, I think I think that's actually such a good point that I can't believe we haven't really got into until now.

But like when we talk about the tools that AI can use, we're still just talking basically about software at this point, right? Like everything we've covered is AI agents only really use software. When you bridge that gap to the physical, that's when some of the scary stuff can happen because yeah, there's a ton of stuff that you could automate that is lowrisk, like perfectly suitable to this stuff. Um, but yeah, once it becomes economical and the cost of of giving AI agents bodies, like physical bodies, then yeah, that's where that's where the S will hit the F for sure.

What was the quote from Colin that I really enjoyed? It was like um if you think the size of the market for smart cars, for autonomous vehicles is big, wait till you see the size of the market for human labor. And I was like, it's getting there. Right now we produce 50 to 60 million I think autonomous vehicles a year. You know when we're getting to the scale where we can start to produce humanoid robots to that level like it's going to our as society is going to completely change and how we interact with technology how we interact with one another is going to completely change.

But we've been putting off for a long time now. We are going to do some episodes on robotics. I've been finally convinced Jack because obviously I love this topic so much. So yeah we're going to do a few episodes on robotics and talk about say the impacts and what's going on and hopefully get some leading experts. I'm trying to line up to speak about it as well. Yeah. So, tune in in the next few weeks to discover which of Alec or I has been a robot this entire time.

Answers on a postcard votes, please. Thank you. So, on that bombshell, yeah, we'll say thank you for listening. You can join us next time to untangle a little more web

Topics: AI & Robotics · Blockchain & Crypto

Related episodes

All episodes