Time to decode the buzzterm du jour - in this episode of Deep Learning, Damien and Nikola discuss the emerging (or is it?) concept of agentic AI, and how AI systems that act autonomously and perform complex tasks could transform individual and corporate life. From planning vacations to automating business processes, we talk about the efficiency gains and potential risks associated with Agentic AI. We also touch on its implications for the future of AI technology and industry adoption. Don't miss insights into how this evolution could shape an AI-first world. Follow PolyAI on LinkedIn Watch this and other episodes of the Deep Learning pod on YouTube
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
NikolaI think well over half of all trades are like non automatically initiated algorithms, right? Right. And of course, it may be why we end up with market crashes that we sometimes don't understand, or they could be exacerbated. Potentially, right? But equally there are a lot of people who think that agents doing things online will be well over half the web traffic very quickly in two to three years from now. So, you know, we could get there even faster.
DamienHello, and welcome back to another episode of deep learning with Polly. AI, the only podcast that successfully cooked a 22 pound Thanksgiving turkey a week ahead of time. Well, I'm just kind of bragging for myself because I did have a friend's giving this past weekend. Do
Nikolayou have a bigger one coming up or
Damienno, I'm done. I did. It's a one and done thing for us. We go elsewhere for the old Thanksgiving feast. But what about you, Nicola? Do you celebrate the American holidays at all over in the UK?
NikolaMy wife was born in the U. S. So we have a healthy amount of U. S. cultural exposure. Plus, like I tend to work predominantly with Americans. It is awesome because my calendar looks like someone nuked it, right? There's nothing in it. I am very, very looking forward to Thanksgiving. And unlike you, I don't have to cook a turkey or anything. Maybe I'll deliver one,
Damienyeah. Why not? Why not? I'm glad that you're thankful for an empty calendar and that's for
Nikolaan empty calendar. Finally. Yeah.
DamienWe'll send you and everyone who's listening or watching into the holidays on a good note. Actually they'll probably be seeing this when they get back. So hopefully you're nice and full. Maybe you've notched a little bit up on the belt. But you're coming back and ready to learn about. What everyone is talking about, within the industry, which is this concept of agentic AI. So that's today's theme. Nicola, when you initially hear agentic AI, how do you feel?
NikolaI think we're just like adding new terms to explain something that we all like already understand, right? I think it's like allowing yourself to think about you know Just like what you'll be able to do with LLMs going forward, right? So it's like a big umbrella term for like things that will Actually take actions in the world that will take sequential things that will call other things that will have these like chains of, like calling other software, other agents to do things on their behalf and things in parallel and things in the background. It's really like an excuse to read to let your imagination run loose.
NikolaAnd. Let me stop with that. But I think, how do you see it?
DamienYeah, to me it's funny because we're obviously in the industry and we see, or the concept is, is that we're producing voice assistants that are essentially virtual agents, what other people would refer to, but the agent in agentic AI is referring to agency is referring to a step towards more autonomy, so less human input into this new wave of artificial intelligence. And more capability to act unto itself. Right? So traditionally I is about computation. It's about probabilistic outcomes, right?
DamienGenerative AI is about actually creating something agentic AI. To me is all right now, let's act instead of just creating output. Let's act on this. And so it's my understanding that agentic AI started on a, on an individual consumer level, right? Having artificial intelligence represent me and my actions, going out, booking airline tickets for me, setting reservations for me, et cetera. But obviously there's a lot of corporate applications to this as well. What do you see that as being the migration of this concept of agentic AI is how can we take this representation of an individual.
DamienAnd it's a technological butler and putting it in the corporate sphere.
NikolaSo I think there are many definitions that people say that it's like AI can act autonomously that can. Perform things independently. I think that it really is defined in contrast with a co pilot approach and the assistive approach where AI is there to help you be better to, find things more easily or inform your actions as opposed to performing actions on your behalf, right? I think in that. What we've always done is voice agents, right? Things that, with every enterprise client, we look at how we can do work on behalf of not on behalf of the agent, but instead of the agent, right?
NikolaLike make reservations make appointments re deliver things, change your account details, navigate your subscriptions, like change, update them. So all that stuff has always been about. What we today call agentic, right? The copilot pieces where I think a lot of PR went first, right? So I think this, the second piece, of course, is like the whole thing around can they improve themselves, et cetera, that's overloaded and doesn't have a clear definition either, right? Is that
Damiena necessary function of what people consider agentic AI is the ability to self improve?
NikolaWell, I don't know. I think you, if you Google in three different ways, you'll see. Yeah,
Damientrue.
NikolaAdmissions. People talk about that. They talk about like productivity, they call about, it's, there's a lot of stuff. So, no, I don't think it's a requirement. I think one thing that was always something you could imagine as something that runs in the background and performs many actions. While, it does the work. It's it has agency in their life. I'm like, plan a holiday for my family and then book it and put it in the calendar and move my other things out of the calendar and book holiday, perform like an ordered sequence of actions, right? And potentially, call a lot of APIs, figure out how to do things.
NikolaIt really isn't, I think, strictly defined, right? It is about just Doing work doing a lot of work like potentially something that's not like click. Here's a response or click, right? But something I might go and be a workhorse right and that kind of like you're imagining this High powered sidekick that maybe emerges occasionally to clarify something. But other than that, it's like an autonomous Entity doing work. Right.
DamienSo going to your example though, of planning a trip. So right now people are, who are just getting acclimated to the concept of generative AI and LLMs out there assisting with, taking a step above doing basic search, right? So I search, tell me all about Turin. Okay. Here's some information. Okay. Now, GPDs of the world and they're saying, how about. You tell, or I guess they're dropping GPT right now. It's just chat, right? Okay. So they're going to these LLMs and saying, plan an itinerary for me. Here are my interests. And then you get some text based output.
DamienNow it's to your point, tell me about my interests. And then also start booking like travel, start booking.
NikolaPlan an itinerary for. A trip that might cost under 3, 000 pounds for two adults and two kids.
DamienOh, sure. Yeah.
NikolaI'd like it to be, like in the South of Spain or Greece. And I'd like it to be, I don't know, all inclusive because we have two small kids and we can't really autonomously. We can't, we can't be very agentic ourselves. We're a vassal state to to the kids. So it's that thing might go. Check the prices, right? Not just like general prices based on like
Damientrue or
Nikolanot, but like, Hey, I found this, this, this would you like me to make a booking? Here's a proposal hit? Yes. Right. Or you can go like, no, no, no, don't do that place. I've been there. But I like this thing in Greece tell me about that and maybe, check flights, check which days I might have to, can I like tweak dates a little to save money? Can I fly out on a Sunday instead of a Saturday and take an extra day of holiday here, right? Like and then that thing could go and just like query a lot and that was always the dream of the superhuman piece of it, right?
NikolaYou know how hard it is when you're looking like this, that, this, that, and you get annoyed. You, did you look enough or did you just end up like booking or did you really optimize it? Who knows, right? Because it's often like a, like it's a, there's a Cartesian product to the whole thing. Like when do you fly? What are the accommodation dates? You need to go and start making a spreadsheet, right? Which you may or may not do. So this thing should be able to just present you with those choices. And then you go Oh, well, this one,
Damienit seems like it alleviates that analysis paralysis that you get which is that when you're inundated with different options that traditional like output from LLMs is going to give you, now you can just say, look, Do it and then book it because you're making things more stressful by giving me too many things to look at. And I need to now take the perfect itinerary and put it in action or something.
NikolaIf you know that you can change the dates and check, and then you're looking at you may have different like thresholds and you'd like to fly with, I don't know, United, but maybe you're willing to not get those points if it's 500 less. Right. Yeah, sure. You could just be like giving different options with One world versus star alliance things, but just check it all for me. Mindlessly, go check it all your AI. You have time. Right. It comes back to potentially very strong ROI, right? We'll ignore the cost of compute. So we, yeah, seriously, in Brazil has evaporated as you tried to save those.
DamienYou just burned three grand and took out a 17 acres of rainforest. Okay, so we see the personal application. Obviously, there's a lot of application as well for this concept, this general principle and these aggressive models that are working toward that end goal within the enterprise space. The thing that always strikes me, though, is that we, as individuals are a little more willing to absorb risk when we're, evaluating different, Flight plans or whatever. And, we're putting in prompts and we're refining results accordingly, then it's willing to take the act, but we're still in a society and a global economy that is walking towards generative AI a little hesitantly.
DamienThere's been broader acceptance, but there's still the question of risk. How, why are enterprises investigating this, what I would call next evolution of generative AI? When they're already a little hesitant on simply getting input or providing output from generative AI to either customer base or internally to improve operations. Now they're basically saying, okay, we're not only willing to do that, but let them act on our behalf as well. It seems a little cart before the horse for me. What's your opinion on that?
NikolaI think the expectations and risk appetizers are still settling, but I think there's a huge variance of them. I think you'd see a lot more risk appetite if we saw more of these like attempts come to fruition a bit more quickly I don't think that they're like companies that are like willing to let it do all this and then like basic things like there is I think A funnel right where there are fewer and fewer that would allow it to run amok and make these decisions It doesn't mean that it's making decisions if it's agentic It might mean that it's just performing a more complex set of actions You To help someone do something.
NikolaI think at the end of the day, it's a business process that you're, or you're deciding whether something works or not. And I think, like with any other technology, you're going to have to test, be like thoughtful around how you've evaluated it when you continue to use it and when you don't, you know, think of all the hedge funds using. Like algorithms for trading, right?
DamienTrue.
NikolaThat's agentic, right? It may not have been, LLN powered agentic stuff. It may be powered with a far more sophisticated machine learning model,
Damienright? Yeah.
NikolaThat does mathematics that is a lot more complicated than what an LLN does for now. That's okay. We use these systems, right? And we just have Ways we have like off ramps that say, if this starts happening, we'll stop. And you have continuous measures around is it producing value? Are we, what's the P and L for the day, right? Where do you like stop trading? Similarly, I think, if this thing is like doing part of the work of humans, but you'll be measuring just I don't know if it's these reservations and things, I'm no expert on that, but you'd probably be looking Just has something changed drastically.
NikolaIf it has, maybe there's a bug that you haven't planned for.
DamienWould you say that, going back to you commenting on, ag agentic, is it just a bit of a buzzword? It is a little a pr at work? Would you consider APIs to be a, a rudimentary form of ag agentic action API is
Nikolaabsolutely like agen and that, it's,
Damienso this has been happening all the time. It's just all of a sudden, maybe it's the ll empowered ness behind it that just accentuates. I think one
Nikolathing that is exciting, that gets I think a lot of people excited is, you saw an LLM and its ability to just look. It has consumed, like, all information that it could get its hands on, right? Like it can probably tell you quite effectively about you and me, no problem, right? It can tell you, the extended history of your town, like everything in it, and maybe well known, like hotels and landmarks, and it will just know things, right? But what if it could, and this was like the first thing, like when 3. 5, like in public reception became what it was, you had people like doing these API things where, they would instruct the LLM to Return the bits of information it doesn't know about so that you make a function call you go to google You pull down some information you reinsert it into the prompt and you return it that way, right?
NikolaNow we have that like like in built with searching summarizing etc I think that was the first like outlet of what we kind of hope, or I don't know if we hope, to some extent we hope and anticipate that it becomes the CAGI in a way, a pseudo conscious thing that can continue to exist, it has consciousness and continues to interact with the world, right? For it to interact with the world, it has to be agentic and that it has to have these ways of interacting with the world. It has to be able to at least search the internet and read things on the internet.
NikolaIt should be able to call API calls and, potentially interact with other analytics, right? So that it can go and just do a lot more and maybe learn from those actions, right? And get continuous feedback from the world around those actions. If I let that system be like a control system for temperature, I've been in this room boiling all day. And that thing could, observe that from certain things. If it was the controller and go okay. If I increase the temperature, Nicholas, heart rate and general health, I don't know if I had an Apple watch or whatever. Something's off. What if I would cool him off? It could experiment, right?
NikolaLike you could do all sorts of things to make my life better. That's a bit of an abstract example, right?
DamienI think that also gets to the application of. When we see AI represented in a negative fashion in modern pop culture, it's a lot of increased agency leads to overabuse of power. So there's been horror movies that have come out that have shown like what happens when AI gets too presumptive and of course, even throwing back to the ultimate Skynet like singularity. I think that there's easily something that is is triggered in the individual's mind when it comes to. I'll use you as an information resource. I'll use you as an ability to endure massive amounts of compute to give me answers that are helpful for me.
DamienBut am I going to let you start making decisions on my behalf? So it seems like the autonomous aspect is both welcome and interesting and Stokes, a lot of like interesting curiosity, but it's also. A massive trigger when it comes to how much do we let these systems act on our behalf, you did mention about how this was like steps towards AGI, and so it seems like that's the natural progression where do we see it, foregoing risk with regards to enterprise adoption, where do we see that the desire to, to capitalize on on AI's ability to act in an agentic application.
DamienRisk aside, is this something that you think ultimately drives more benefit than than that added risk? Maybe not risk aside, but risk in mind, do you think it drives added benefit for standard enterprise operations where it could be applied?
NikolaOf course. Look, it's when you think of a genetic behavior like what you would want to think is that there's one thing thinking about like your request, and what do you, as a city of cold, you asked about something, and then maybe the second AI is there to kind of like. Not demand APIs, but instead take over the screen and go through like the pages that a human would click through, et cetera. And I think that's what a lot of people imagine. And by the way, that thing then would not be, an API might not be instant either. But it's like they're waiting, making progress, learning. Maybe it can clone itself a thousand times and do the work a thousand times in parallel.
NikolaThat's something that a human cannot do. That's how you can get your 1000 quotes, look at all of them and come back. Maybe. Even if you don't have an API to do it, right? Like it is able to just do these things and come back to you. With the best solution, right? It could work on your behalf that would take a human, let's say it was getting a thousand quotes Like in each quote took a human two minutes to fill out click go through some right look at it put it in excel spreadsheet. This thing would just do it all at once, right? And maybe without any extra apis, right? Maybe it's just they're you know, like rpa companies like controlling the mouse and going in like doing the work, right?
NikolaThat would save companies a lot of money on that digital transformation, right? So You Rather than having to, potentially, and by the way, if it's of sufficient capability, it should be able, maybe, to over time write that API and go look, I don't want to put it back a thousand times I'm going to go and see, all right, this is the architecture, I'm going to say, realize that this, I need this function here, I might send an email, To another agent who's like a proxy for a software architect in that, in the other part of the company who has clear authority to give me access to this or that, right?
NikolaLike I'll ask them to produce like to open a certain port to a certain IP address. And if I send. And like an HTTPS request of this kind with this like key it can send me all that back. And maybe that goes huh, Roger that. Okay. I'll go and check. And your agent like snoozes, whatever waits for it. The other thing might surface it to a human for final approval. I'd say, Hey, I'm okay to open To this kind of function, that port that, that request from this, you might even have two humans in the background going Hey, Mike, did you,
Damienautonomy, but not that much
Nikolaautonomy. You ask for this is this something, yeah, it's not, that should be explained. They go, okay. Okay. Yes. It says yes to its agent. And then it's okay the two phone operators put them through write some technology and do it anymore because, I remember the first time we had Google Duplex call one of our assistants. And they had a clumsy conversation about opening hours of of a pub in the UK. It was like 2020. It was hilarious. We still get those calls happening daily, right? Because that's how they update Google Maps opening times.
NikolaAnd it was like two AIs who were speaking clumsy human language. Ours has gotten a lot better. I think theirs might have gotten a bit better, I don't know. But that's gigantic, right? Their agent calling the agent of this thing, both with delegator work. One agent was delegator work that said, hey, pick up the phone, help people out. The other one was like, hey, here's a list of restaurants, go call them. And these two agents were interacting. And doing the work and like publishing results online. Isn't that amazing?
DamienIt is. And that, that Hey,
Nikolalike you get this question all the time. And then they ask if it's like a two part thing, there's a break in the middle. Can they deliver things in the middle? We surface that to that client and They can change their response so that the next time the decision calls, the information is changed.
DamienSure. There's, there is something that's particularly interesting around the observability of agentic AI interactions, especially two way agentic interactions, because as we talked about earlier, this has been happening all the time, behind the scenes,
NikolaI think well over half of all trades are like non automatically initiated algorithms, right? Like when it comes to a global economy, it's stability and everything like we should be comfortable with a lot of other things. And of course, it may be why we end up with market crashes that we sometimes don't understand, or they could be exacerbated. Potentially, right? But equally there are a lot of people who think that agents doing things online will be well over half the web traffic very quickly in two to three years from now. We could get there even faster.
NikolaAs long as we find good use cases where it makes sense to burn all that compute.
DamienThe rating was on the wall leading up to this. Go to your local grocery store and half the people that are in the aisles are shopping on behalf of others who have booked them through an app. So again, observable, agentic, non AI, but. Still agentic. Nonetheless, we're moving to an agentic economy and the market is responding. As a matter of fact, some former just reported today, I think, or yesterday that dev slash agents, which is a startup that was founded by some X meta Google stripe execs just came out with a 56 million seed round raise. It's an O S for.
DamienAgentic AI. So is that going to be the next big wave of startups are in our companies that are basically creating products for agentic AI? I
Nikoladefinitely. As you said, I think like the naming shifts, it's been happening all along, right? Kind of comment on that one. I don't know much about it, but yeah, it's not that
Damienparticular one, but just in general, is this, is it in the steps towards AGI and how
Nikolawe have agreed architectures and standards on the internet, I think we'll have them in, the internet post, like an AI first world, right?
DamienYeah, true. All right. No, that makes a ton of sense. Interesting. Any other like final comments, thoughts on the future of agentic AI or where this is, this just again, another step towards the ultimate end goal of AI first world placement. But
NikolaI don't know. Look, I think like we, let's go build some cool shit and see where we end up.
DamienThat's the best way to outro. All right, cool. Nicola, appreciate your time. This is good stuff. I will be looking at every single piece of automation that I have around me, whether in the behind the scenes operationally, or in front of me, observable agentic AI, as we were talking about with a whole new,
Nikolagood luck going about your day, man.
DamienSame to you. And for the rest of you, we'll talk to you next week. Thanks, Nicola. Take care.
Three years ago, Simplyhealth's PolyAI deployment was only resolving simple FAQs over the phone. Today, they run voice, in-app, email, chat, and WhatsApp conversations on the platform. Nikola Mrkšić talks with Paul...
An agent that gets better every day compounds into a very different agent within months. That's the idea behind Wren, PolyAI's new release: an agent whose job is to make your other agents better. Some of what it finds...
Companies are pouring money into AI, but not everyone is getting the returns. Ian Jacobs, VP and Lead Analyst at Opus Research, thinks he knows why. In this episode, your host Nikola Mrkšić sits down with Ian to work...
When PolyAI built its Builder MCP, one client helped shape it: Vixxo Facility Solutions, one of the largest facilities providers in the US. In this episode, guest host Damian Sasso talks with Vixxo's CTO Derek Neighbors...
PolyAI just launched Dialog-RSN-1, its first audio-native model, and your host Nikola Mrkšić sat down with its builder, Matt Henderson, to unpack why it’s a game-changer for building voice agents. Most voice AI either...