The Agent Exchange: Practitioner Insights

Agents in Production [Podcast Limited Series] - Episode Five, Dmitri Jarnikov, Chiara Caratelli, and Steven Vester join Demetrios to explore AI agents in e-commerce. They discuss the trade-offs between generic and specialized agents, with Dmitri noting the need for a balance between scalability and precision. Chiara highlights how agents can dynamically blend both approaches, while Steven predicts specialized agents will dominate initially before trust in generic agents grows. The panel also examines how e-commerce platforms may resist but eventually collaborate with AI agents. Trust remains a key factor in adoption, with opportunities emerging for new agent-driven business models.

Demetrios [00:00:00]: Buckle up and strap in. I grabbed three friends to talk about two topics related to AI agents in the e commerce space for the great practitioners. Discussion Number one, are we heading towards a world of generic agents or specialized agents? And number two, what is the future of companies making shopping agents? Let me introduce the three awesome guests I brought in. Steven from OLX who is the head of product there working on Gen AI efforts. He's focused on building AI enabled solutions and doing a great job, some of which you may have heard of from our last AI Agents in Production virtual conference like OLX Magic. I've also brought in a friend of the pod, Chiara, who's back for a second appearance. She's a data scientist on the process AI team. And if you haven't listened to the episode where we dove into web agents, we talk about all the work she's been doing testing out different frameworks and eventually building out their own web agent framework for the process AI efforts.

Demetrios [00:01:17]: And last but not least, the most opinionated man in AI, Mr. Dimitri, the Senior Director of Process AI. This man's a deep thinker. For the past few years I've had the pleasure of being able to converse with him about all topics AI and he's been working on gen AI products, including them good old agents. It's always nice when I get to learn from him. As a reminder, we are using a simple definition of AI agents, which is AI that can reason, execute tasks on a user's behalf and use tools. So a very simple TLDR way of putting it is AI that can break out of the chat box. Let's get into our first topic of the day.

Demetrios [00:02:09]: I want to bring in our guests and hear their hot takes. So generic versus specialized agents. To Dmitri, you were introduced as the most opinionated man in show business. I'll let you kick it off for what you think the future holds when it comes to generic or specialized.

Dmitri Jarnikov [00:02:30]: Well Dmitri, thank you very much for the introduction and also for giving me the pleasure to be the first person to provide the opinion. Well, I think when we talk about agent and I like very much the way that you form frame the discussion, first of all we need to actually look at what type of agents are being built, right? Because in the past couple of months we see a lot of companies, group research teams putting out either commercial products or even open source frameworks to provide functionality for people to build agents. And roughly, at least in my opinion, we can group them into a number of categories. So the first category is what I call generic agents. So on Extreme case, these are agents that are not specialized in terms of reasoning about the task and also not specialized in terms of execution of a task. So think about computer use, for example, an agent that can take over your computer and just do whatever you ask it to do. Think about web agent that can go online, use your browser to either shop for flight, shop for an item, or just do some kind of web research. So these are agents that most of the time are not specialized.

Dmitri Jarnikov [00:04:00]: And then on another side of the spectrum you've got agents that can do only one task and only on one website or on one service. Sometimes they are so specialized that they require a specific API to be able to execute the task. So when I look at this, I think that, well, specialized agents in general are the most precise in doing what they require to do. But business wise, this is not scalable approach. Imagine that you are building a solution that needs to shop online. You cannot really build an agent or a family of agent that can go and be integrated with every shopping site in the world, right? So it just doesn't scale. On another side of the spectrum, if you've got very generic agent, so what we see, they are good at everything a little bit, which means that they are good at nothing in particular. So they can do a task which requires couple of steps of execution, the most common that we've seen or we've heard of, they can do up to six steps in the sequence before they hit the wall.

Dmitri Jarnikov [00:05:26]: And if your task fall within the short execution pattern, you're fine. But if you've got a complex task that requires 10, 20 steps, these generic agents, they cannot just achieve a result. So when I look at agent, I'm looking at two dimensions of implementation. I'm looking at reasoning capabilities and I'm looking at execution capabilities. And in each domain I'm looking for specialization. So in recent capabilities, is it something that is not specific at all? Is it something that's somewhere specific? So I'm thinking about food ordering agent, right? So it's an agent which understand that to order food you need to understand preferences of people. You actually need to understand people. You need to build what is called food ticket.

Dmitri Jarnikov [00:06:19]: And then you need to go online, for example, and visit a couple of services providing food to make an order. And then on execution part I'm looking at again agents that can just use your computer or use your browsers. And being absolutely not specialized to agents that can do well on specific websites. So if you look at the latest product that we saw, we saw operator from OpenAI so in my opinion this is an example of agent that is fairly generic in reasoning part and has a combination of generic execution with specialized execution. And my opinion that the future is for agents that are reasonably specialized in reasoning and have a combination of specialization and execution with the possibility to fall back on generic execution pattern.

Demetrios [00:07:25]: Bit of a hot take there at the end when it came to the specialization and reasoning, but got to be honest with you man, I was hoping for some stronger opinions. That's okay though, it's just a warm up session. I get it. We've also got Kiara to take us through some of this because Chiara, you've been working with the generic form of these agents for a while. What are your thoughts?

Chiara Caratelli [00:07:47]: Yeah, so actually I started working on generic form of web agent but then quickly deviated towards some more specialized agents at the end a good balance was something in between like Dmitri just said, because the powerful thing of agents is that there is not a fixed workflow they can adopt based on what's in front of them. Right? Kind of like a person would. And agents can call tools that can be agents themselves. So you can have very specialized agents that are still managed by a generic agent and vice versa. So one thing could be having a specialized agent for shopping that can browse platforms, knows how to browse bigger platforms very well, but still can be generic web browser when it comes to smaller websites that it has never seen. So with agents it's much easier to have this kind of mixed approach compared to previously. I'm thinking this is a discussion generic versus specialist that predates agents. It was there with smaller machine learning models versus bigger models.

Chiara Caratelli [00:09:02]: It's always been around. But I think agents broke this pattern because it's much, much easier to dynamically choose what's best. And we see it also with recent tools that came out like OpenAI operator, they made a call for action for websites to integrate with them so they prefer to use APIs to integrate and do things fast while they also can browse the web in general. And I think from point of view of customer interaction, the biggest changes in speed. Of course there is success rate. It's much better for specialized agents if they have their specialized scenario in front of them. That's a fact. But if you have a long enough, a long enough time and number of calls you can do like with reasoning models, you could still get there.

Chiara Caratelli [00:10:03]: The problem is if you're interacting with a person in real time, this is not possible. So this is why specialized agents are still very relevant. If you have something like Deep research that. Well, we saw it with Gemini OpenAI also recently released this. Then you can have a generalist web browsing tool that can be slow. They don't care how much it takes, but it gets to the answer and the user is happy at the end.

Demetrios [00:10:38]: Speaking of users, Steven, let's bring in your product mindset. I know you like to think about how users are going to react to these new tools they get. Where do you see the world of generalized versus specialized going?

Steven Vester [00:10:53]: That's right. So yeah, let me bring in the user angle because I think I wouldn't be doing my product background justice if I didn't. Look, the bottom line is I don't believe in sort of 0 to 1 full adoption straight away. I think it would be very iterative and I think a good way to think through this, or at least the way I think through this, is with an analogy of a human assistant. So imagine that you get an assistant today and imagine they're just finished uni and they're going to help you with whatever. Probably in the start you will just ask them with, okay, maybe get me a coffee. Hey, maybe here's some raw notes I wrote. Maybe predefy it, maybe use ChatGPT to do so.

Steven Vester [00:11:38]: But essentially you'll give very small chunks of tasks to them and then I will trust, without knowing their capabilities, I will trust them to do that small thing well, like getting me that coffee. Then over time, as my trust builds, I might dare to give them a bit more complex tasks. So I might ask them to write a certain report or research a certain topic, or create a plan for a certain area and I will change very much. How much context I give to that person, I will change. And I will also do it only when I have a proven track record of their capabilities and my trust in them grows. And I expect this analogy of how we would use a human assistant in our lives to actually work very well for how we're going to be adopting agents and assistants in our shopping and throughout our entire life anyway. So that means we'll start off with a very small specific purchase and we will see how it does there. For example, buy me movie tickets to movie X on Saturday night in this cinema.

Steven Vester [00:12:47]: So there's not much to mess up there. Essentially my trust grows and capabilities grow. I might do more complex stuff. Hey, buy me a new snowboard. Now that framework of actual abilities versus perceived abilities, I think those are the two dimensions that are going to be needed for people to trust more and more in agents and for agents to Cover more and more parts of customer journeys and for users to accept that. So the truth today, and this is what Chiara and also Dmitri already say today, generic agents are just not good enough to really do stuff. There's errors, there's. It will make a bit dumb mistakes here and there.

Steven Vester [00:13:32]: So today you need a specialized agent, one that is catered to one niece or use case or one particular platform and they're actually going to be able to do certain things. And so I suspect if we follow that framework of actual capability development and then my perception of that capability development, we're going to see the first real success cases with those specialized agents. So customer is going to be trying those and adopting those and building some confidence with those. And so you're going to start to build some trust with that agent in that domain. And just because the capabilities aren't there, I think specialize is the only way today. Now thinking long term, I don't think there's a moat here. I think that capabilities are a short term problem and our perception of capabilities are also a short term problem. You know we have O3 now, some great stuff coming out with Deep Seq.

Steven Vester [00:14:29]: There's obviously no sign of stopping here. And so where are we several years from now? Well, I think several years from now generic agents like computer use, like operator, they're going to be fantastic, right? And in that world, I don't know if I still need this generic agent that I've come to appreciate and trust. I could imagine just like you use certain Google tools already now, it's just common, of course you're going to use it and I'm sure certainly Google is going to launch one and more labs will as well. So I would expect that over the longer term customers going to be moving to generic ones as their capabilities improve. And maybe one more analogy here. I don't know if you remember how you started using online payments and mobile payments, but it feels very similar to me. You probably did a very small payment somewhere and then you started doing it in more places and you did it first on those very specific checkout flows. There was no generic one.

Steven Vester [00:15:26]: But now you don't care, right? It's all just common. And you have maybe your PayPal or you have your Google Pay or whatever you use. I suspect a similar type of curve there. So bring it back to the discussion. Generics versus specialized. I think today it needs to be specialized because generic is not good enough. And I think that is also the only way users are going to be adopting this technology is through These good experiences small and then bigger, bigger, bigger with specialized ones. And as the generic ones get better, there will be a move there and then it's just going to be as common as online payments is today with some big names that you use.

Demetrios [00:16:03]: Okay? I like the idea of perceived capabilities and real capabilities and that delta between the two. I think that's one of the biggest disconnects that we have right now in the AI world. We see so many different demos around the Internet and it makes us say, whoa, look at that. What can AI do? It's incredible. And then we go and try and play with it and it falls flat on its face.

Dmitri Jarnikov [00:16:31]: I actually think that Steve made a very important point, right? And especially what I like about his argument is this historical perspective. I am coming from slightly different technical background. Many, many years ago, I was working in IoT space, right? And we also had this discussion, okay, what the future will be in IoT in terms of network protocols. You remember those? Would it be IP like for the rest of Internet, or would it be specialized protocols? And everyone in the field was saying, well, you cannot do full IP stack on those small devices and you need to be really specialized. You need to be real time, you need to be light, you need to be this and this and this, because otherwise it doesn't work. What people forget about is what Stephen actually suggests to think about temporal dimension, right? So yes, we cannot do many things now, but probably as the most widely adopted, most generic solution will win because yes, LLM will get better. So it will better get better in reasoning. So yes, it will be able to say, okay, I've got, I can cover food ordering, I can cover buying items, I can cover buying trips or whatever.

Dmitri Jarnikov [00:17:54]: So it will be better in reasoning over time. And I think it will be better in instructing execution part of execution agent or execution part of an agent to go online and do something. So there is, I agree, very high likelihood that in the future we'll just go for generic, but for some time it will be indeed very much specialized agents. Another aspect that I want to probably mention is that part of this discussion is actually not technical. You know, where technical people. Stephen, sorry, I bring you into this group of technical people as well. But as a technical people, we have a bit of myopic view on technology because we say, okay, this is what is possible, this is what is reasonable. What I'm also looking at when talking about agents in E commerce, what is the dynamic between companies making agents and companies making websites or platforms, right? So you can say, okay, Generic agent will be able to do all the things on the platform, but most likely it will not be able to do it effectively.

Dmitri Jarnikov [00:19:15]: Right. So think about it's agent which will mimic how we behave on the platform. I don't know what's the last time you did shopping? I'm absolutely sure I did not use an optimal path on Amazon or on a booking or any other website. Right? So I clicked around like a monkey. I look at things that probably were not relevant to the end result and now imagine agents doing exactly that. So specialization will be still valuable for platforms.

Demetrios [00:19:51]: Let's take something into account though. Companies make a lot of money on those random clicks around the website because I myself have gone to a website countless times with the intention of buying only one thing and being enticed into buying a few extra items because of those damn recommender systems that know too much about me and my vices. So how do you cope with that in a world where agents just go and get you the thing and they in theory are not susceptible to recommender systems?

Steven Vester [00:20:27]: I did think about that and the answer is that it's hard. Look. So the ideal agent can execute a, let's say a buying agent, right? Shopping agent can help you outline the right criteria for what you're looking to buy and then find every possible match. Right? So recommenders are irrelevant. Ideally this agent can go through the entire inventory on a certain platform and just find that perfect one. And actually not just one platform could be multiple, right? If we talk about generic ones. So this basically means that the better access that this agent has to the entire inventory, the better job they can do finding, retrieving those things. So it's probably not possible for that browsing agent to just go click page by page by page throughout all these millions of listings in some cases.

Steven Vester [00:21:25]: So there will need to be some kind of other, like a middle layer here. That could mean if it's an external or even a bit, let's say a bit of competitor or threatening way you could see it like maybe they would have to scrape all your inventory all the time, have that in a database and then make that database searchable. So basically scrape inventory and create some kind of quadrant, set some quadrant database and have the agent look through there. But another way could be that platform actually opens that. So you open your doors for agents and you make it easy for agents to find the right thing. So recommender systems, they become different. They're not showing you what is recommended, but rather gives you access to all the embeddings that have been calculated, stored in some vector database, and you just say, okay, here it is, look what you need. So, yeah, don't have a strict answer on where things will go, but just.

Chiara Caratelli [00:22:24]: Some thoughts if I may add to that. I think platforms could even use their own recommender model for that, so they could decide to expose an API where the results that they propose are already ranked by their own engine. And I believe that could even be a solution for ads in that regard because they could add content that is boosted in that actual response that the agent could use and then present to the customer. So I think there are a lot of ways here and it will be really important to see how the user interacts with all these new modalities. And to add to the previous point, there is this kind of conflict between generalist agent and specialized agents. But if specialized agents open new use cases and we see adoption through, through those use cases, this might be that these specialized agents become the generalist. And on the other hand, generalist agents can open the door for specialized agents and adding new modalities, new tools and so on. So, yeah, I think customers at the center of all this and end users would be really important.

Demetrios [00:23:49]: Yeah. Okay. I want to highlight two things you said which I have been thinking about at length. One is the recommendations that come back to you in the answers that are fed to the user. I could see that being something that you get when your task is completed. And then there's another part that says, do you want me to do any of these other tasks? That is becoming a fairly common UX design pattern we're seeing with agents these days. So it's not that big of a jump to think that that's going to be next. Then the second part that you said, Chiara, which I had not thought about, is the generalized agent being general at the beginning of the flow, and then it kicks it off onto a specialized agent to accomplish the task.

Demetrios [00:24:40]: Doesn't feel like that far of a stretch because there are some design patterns with agents that that's happening in. So once maybe we build out a strong network of specialized agents, the generalized agent is just there to do air traffic control in a way.

Steven Vester [00:24:56]: I think we'll enter a world where you want to build tools that agents can use. Right. So you want to. And in that way you could think of a specialized agent internally built on a certain platform, just offering its capabilities in a sort of API to another agent, a generic one like let's say Google's one or whoever will build it, and then you get a Bit of best of both worlds thing. But yeah, I think the problem there is not so much on capabilities, but it's rather a business strategy kind of thing. Do you want that? Right. And I think that's a good bridge to the next topic as well.

Chiara Caratelli [00:25:39]: To add what Stephen said, we've been experimenting a lot now with LLMs being used as recommender systems. So an LLM that knows the user, knows what they have done and can put rationale inside all this behavior can actually provide input to another LLM during a recommendation. So there are already agents that work together with other agents to provide a better experience for both user and platform.

Demetrios [00:26:10]: Since we are already kind of drifting into this space, we might as well talk about the second part of this discussion which is around the questions of what is the future of companies making shopping agents. And Steven, you brought in a few strong points, so I'm going to let you kick it off because really there are so many different parts to this. Where do you want to start?

Steven Vester [00:26:37]: First, I am probably not the best to understand or explain what those companies building those agents will do, but I think I am well positioned to answer what existing e commerce platforms will do in response. And this will guide then of course what those first companies what they'll be doing. So I suspect that e commerce platforms, and I come from the classifieds marketplace world, so that's used goods, cars, also real estate, jobs. I believe that these platforms will start being very defensive in the beginning and then we'll move to embrace it. So what do I mean? The defensiveness would come from the fact that these agents, especially if it's a third party shopping agent, browsing your platform, it removes your customer connection, right? So it damages your brand loyalty. And these brands have many years of investment behind them, hundreds of millions in some cases. It also makes it hard for these marketplaces to get clear customer signals of what's happening on their platform. Right.

Steven Vester [00:27:45]: As agents start browsing, you need to find some way to identify those and kill that noise. Otherwise it will be very hard to do any customer research or build for real customers. And I think a small analogy here is that of price comparison sites. I think we've seen that typically classified websites don't like to cooperate with those so with those aggregator integrator type players because again, you disintermediate, right? You lose the connection to that customer. So this is why I expect platforms to start being defensive and I think they will take defensive actions to make the platform harder to use for agents, not easier. So what are those? Well, think of the type of things we're doing against scrapers. Right. So you have bot detection.

Steven Vester [00:28:36]: You will look of course at IP addresses where things come from only allowing for example, traffic from residential IPs, not data centers. In anti scraping practices you have things like honey pots which are essentially once you identify a scraper, you don't block them, but you in fact send them to a piece of your database or to a piece of your site that a normal user would never see. You just start generating random data just to overload them and poison essentially the data that they're scraping. They wouldn't know anymore what's good and not these kinds of tactics. And I'm sure there's many more will probably start to be deployed by platforms who don't want agents on their terrain. And I don't think that will work so well. So now my story changes. I think that they will try and I think maybe for a while that's how it will be.

Steven Vester [00:29:26]: But then at some point we will see that agents are here to stay. They will permeate society, lots and lots and lots of users. So then it becomes a kind of you can't beat them, join them. And then I actually expect, yeah, we'll try to make it easier for other agents to be on our platform. So that are those kind of things I just mentioned earlier. So having some APIs open or allowing them to look through our vector store or catalog stuff or whatever, there's multiple ways we could help agents give them tools to be more effective on our platform. You can already see that this is what OpenAI intends to do if they're operator right, they want to have partnerships where they can tune towards those sites. And I think trying to become a partner and trying to help a tool like operator be more efficient is something that platforms will start doing.

Steven Vester [00:30:19]: And this will really mean a different monetization model for classifieds or marketplaces. It means that the traditional ways of just selling eyeballs and views, that's gone. So we need to have other ways. And there are ways, I'm certain, but it will really change quite fundamentally the business model over time. So yeah, I think we'll start off defensive and I think that we'll have some success early but then fail and then I think we're going to change. And when I say we, I mean every single marketplace in the world is going to open up for this because if they don't, they will lose to someone who will. An end state you could think about, you know, a marketplace brings together supply and demand. So if you don't open up, you know, I could just imagine some third party even, like, even decentralized, right? Some, some blockchain database where all offer sits and where all demand can just look for that.

Steven Vester [00:31:16]: That's possible. So you, you better open up and make it easy because that would, you know, give you a right to play.

Demetrios [00:31:22]: Well, Dmitry, I know that we have talked at length about whether or not you expose your data to the Internet or you keep it for yourself to create a very specialized agent for your platform. What are your opinions here?

Dmitri Jarnikov [00:31:39]: Indeed, I think it's a very interesting question. And essentially if you look at what is happening with agents nowadays, and this is also somewhat related to our discussion about generic versus specialized, right? Because generic agent in a way is a stick in the hands of company making them. And the carrot is we can make more specialized agent and we can have an agreement with the platform, right? So the platform has a choice at the moment, which is either collaborate with agents or acknowledge that there will be this generic agent that are indistinguishable from normal users coming to platform and shopping on your platform anyway. And I agree with Stephen, in my experience as well, when I talk to different platforms, the immediate reaction is we're going to block it. But this is, this reaction lasts for like 30 seconds because it's a cat and mouse game and people very quickly understand that, well, you start blocking it, you make life for actual users so much difficult that it becomes non practical. Now one of the analogies I use when I think about this is you remember the story of Spotify and Studios, music studios, right? So you basically have artists, right? And then you've got labels. I call them studio, I probably should have called them labels. And labels, usually their main focus is to find the talent and to basically subscribe them so they become part of a label.

Dmitri Jarnikov [00:33:26]: So if they produce anything of value, it's to be sold by labels. And now Spotify comes and Spotify says, look, we gonna let people listen to what they want and we're gonna completely ignore the way that you package stuff. So we're basically not gonna kind of care about your interest. And to me it sounds very similar to what is happening with shopping agent and platforms at the moment. Because this, the risk are all the same, right? It's what you mentioned. The view of users disappears. So labels do not know what people like to listen to and as a result they cannot do their job well. And also the way that Risk profile presented is very similar because Spotify was, okay, we're gonna go go around you labels and source musicians.

Dmitri Jarnikov [00:34:29]: And now with AI agent, shopping agent, it's very similar because the risk that we see is you've got this agent which goes around the platform and just source goods services from suppliers directly. But what happened? Eventually labels and Spotify struck a deal where Spotify says, look, we're gonna focus on serving user needs and you're gonna focus on making sure that the pipeline of artists does not dry out because this is what you're good at and to actually have mutually beneficial relationship, we're going to share with you usage data so you can make your decisions better. And because now we're going to use all your catalog, we also give you a share in our company. So if we grow, you grow. So they created this codependent relationship which also allows them to focus each on their strengths. So if I look at what is happening now, and especially it was OpenAI operator, I see exactly same dynamic because you've got exactly same carrot and stick. OpenAI says, look, we can use generic web agent on your platform, so we still going to play well, we're going to buy some allow user buy something from you, but we better focus on user experience. So let's work together.

Dmitri Jarnikov [00:36:10]: If you allow our agent on platform, if you do not block us, if you present a optimized version of your website or even give us API access, we give you user data so you do not lose it to your point, right? So you keep to know who your users are and then you can start focusing more of your energy on getting suppliers in, on building this domain intelligence in and allowing us to actually just focus on customizing personalizing experience of users. So in a way for platforms it could be a blessing, right? Because on the platform side you actually now do two things. You basically need to deal with your sellers or ifood with a restaurant or take a lot with suppliers. And at the same time you put a lot of effort in fostering this user relationship, in building user interfaces. So AI agent says, okay, we're going to take this upper part from you, but in exchange you still going to get user data. So this was my analogy. I think I could have presented it somewhat better, to be honest. But this is the thing and actually looking at our companies and I think that we discussed it back and forth already a couple of times.

Dmitri Jarnikov [00:37:35]: I think that yes, we're going to prevent agents from going to our platform. It's like the first reaction, you know, the second most typical reaction, they say we're going to build our own shopping agent. And as a Technical person. I'm like, yes, let's do process E commerce agent because this would be fantastic opportunity for us to apply our technical expertise to have fun and so on. But from business perspective, you just need to ask, okay, so your platform, which game you want to play and if you build your own E commerce agent, for example, how do you maintain one of the core value propositions to users which is multi platform? Right. The idea is if I go to agent, I say I want to buy this bicycle in Poland. I need to be able to go to OLX Poland, I need to be able to go to Allegra who is competitor. I need to be able to go to what Amazon, in case they've got like well priced bicycle and so on.

Dmitri Jarnikov [00:38:47]: So if it's my agent, I'm basically to have a value proposition that open it to competitor businesses. Right?

Demetrios [00:38:56]: Huh. I like this idea of the distinction between sourcing and shopping agents in B2C versus B2B. One thing that you make very clear here is that you're basically creating a new platform that is almost an abstraction of many different shopping platforms. And if you're not careful, you can quickly lose scope of how difficult that is by just thinking, oh, here we're going to create an agent that can go and do everything that 10 different platforms can do. But let's keep this rolling. Kiara, I know you have some strong thoughts, so I want to hear what you have to say about this.

Chiara Caratelli [00:39:37]: Yeah, so actually my opinions are very similar to Stephen and Mistry on this, that now we're in a stage where platforms are trying to block them because it's a big change because I feel like platforms interests and user interests don't always align. So as a user I want to find the best deal possible. Platform may want to present certain items that maybe are not the best deal for the user but are the best deal for them. So yeah, I think this is a bit scary in general for platforms, but as the user will expect more personalization will expect an agent to help them discover products and I think they will kind of expect that the agent will give them the best deal possible in order to trust this agent. We're not yet there users don't trust agent. We're still in the first era of online payments to give the analogy that Stephen gave before. But as more trust will be put into the agent that more expectations will come from customers and I think platform will start to give in on this. There are also more use cases that could be enabled.

Chiara Caratelli [00:41:01]: We've been thinking a lot of these at this because there is a bit of doom scenario. It's very easy to say all the potential drawbacks that agents could have for platforms, they're clear. But there are also new opportunities. For instance, they could automatically integrate with smaller vendors and present the products directly and have an automated pipeline to integrate with their own payment system. So user could buy and search for items on the platform itself using an agent and then purchase the product from another platform. So it will be much easier to work as aggregators here. And the other is sharing customer data. I think in future recommender systems it will be more important to give rationale to really understand the user.

Chiara Caratelli [00:41:55]: Not just having a fine tuned deep learning model on user and items embeddings, but actually having something that can give an explainable answer. So we recommend this product because you've been browsing this and that because you like this type of items, because next month you're going to run a marathon and I need to give you the best shoes. So all these will enable a lot more interactions with customer and customers are going to expand them.

Dmitri Jarnikov [00:42:27]: So there is one thing that you mentioned that I think we did not cover at all and I also, to be honest do not see being covered in general, which is selling agents. Right. So we basically say, okay, shopping agent. So I'm as a user I would like to go and I want to purchase something, but there is so much need also for selling agent. It could be I'm as a user have something I don't need, I want to sell, or I'm as a manufacturer would like to offer this product or I'm as a service provider have this to offer. And my guess, and it's pure speculation and we're not talking about selling agent at the moment because there is no infrastructure, if you will, to connect selling agent with a shopping agent. Right. So this is a part which is missing.

Dmitri Jarnikov [00:43:20]: But yeah, at the moment somebody comes up with this infrastructure, then there will be a gate opening to building selling agent. And then the relationship between selling shopping and the platform in the middle will change once again and we'll have another discussion trying to predict the future.

Demetrios [00:43:46]: Yes, trust is so important and I want to second that. I have always thought about trust from the perspective of if I'm using an agent and I give it my credit card number, am I going to have money in my bank account tomorrow or is it just going to drain it because it doesn't understand the requirements well enough? But here you're talking about how quickly we can lose faith in agents if they're not giving us the best deals. You can see a world where if a buyer is being shown answers that someone paid to get in front of them, the buyer will quickly revert to doing its own research because it doesn't get the best deal. So it's a very fragile situation and how you play that out in the future is going to determine a lot of things.

Steven Vester [00:44:38]: So in our industry, a big part of our customers where we earn most money are with professional car dealers, also professional real estate agents. And it's the same with what you're saying, Dmitry. But some thoughts on that. Yes, you could have an agent that would help them, you know, put all their inventory everywhere. But those tools already exist today. They have tools already and just standard API connections. You know, there's software that allows you to, I think they call it multiple listing systems in our industry. So they have an inventory management system and they can just select like, oh, I want this car on these five platforms.

Steven Vester [00:45:19]: Right. And then the API connection just handles it. Sure. You could automate more pieces. Right. For example, there would be customers calling or sending a message and you could automatically reply to them to make your selling process easier. But in the end they typically want, you know, they want to stay very close to that potential customer because that's how they can sell, that's how they can use their skill. So it's a, it's a very interesting domain.

Steven Vester [00:45:41]: I've been thinking about it whether seller agents in our context. So for example car dealers and real estate agents, if it makes sense. And I'm not certain because I'm not certain what use they will have beyond tools that already exist today. Yeah, maybe it could scale more, you could be on more platforms. But typically there aren't that many classified type platforms in a market. Right. You maybe have like 3 max 4 or so relevant ones. So it's a very interesting one.

Steven Vester [00:46:08]: It might be different indeed for let's say a producer of a good who just wants to sell to every market in the world. They don't care. But when it comes to these kinds of selling agents, I'm not certain how the future will look. Definitely interesting. Maybe it's a topic for another debate.

Demetrios [00:46:26]: Well folks, I think we are coming to the end of the great practitioners discussion. I want to thank Demetri, Steven and Kiara for being here with us today to give their opinions on what the future may look like around these two topics and really being intellectually honest with what is real right now versus what is a little bit more hypey. This was a very special production of the process and MLOps community AI agents in production collaboration it is the fifth episode of the miniseries that we kicked off. I was able to spend a week at the Process AI offices and embed myself in their teams, learn about all the agent use cases that they're working on, and then bring the different experts into the studio with me to explain what they've learned while building out these agents. Since it is something new, I would love if you give me your feedback by either leaving a comment on Spotify or YouTube or dropping in a review on Apple Podcasts or wherever you listen to your podcast because it helps me so much to hear the feedback from you all. If it's too much to leave a review, just DM me on Twitter or LinkedIn or Bluesky. Even hearing your feedback is a great way for me to convince people like the Process AI team that it is worth their time to take the day off and come and chat with me in the studio. Lastly, these awesome people that I've been hanging out with at Prosys are hiring, so we'll leave a link to all their job openings in the show Notes.

Demetrios [00:48:06]: Check it out and peace out everybody.