AI Agents, Crypto, Sovereignty: The Agentic App Store

Recorded remotely.

Hodl Up: Hello, and welcome back to the show. Today we have, and I mean it when I say it, a fascinating guest: Dr. David Minarsch, CEO of Valory and a founding member of the Olas Network. Before we get into the technical side and everything you've been building, David, thank you so much for joining us today.

David: Thanks for having me on. Happy to be here.

Hodl Up: To kick things off, I want to talk about the phrase "AI agent," which has become a massive buzzword in recent months. A lot of people just assume it means a chatbot on your website, but there's a distinct line between an AI tool and a fully autonomous agent. So for people who want to understand where the internet is actually heading: what, in your opinion, is the fundamental difference between the AI copilots we use today and the sovereign agents you're building? And why is it so critical that this next generation of AI is user-owned rather than corporate-controlled?

David: Great question to start with. Let's get right into it. What most people are deeply familiar with by now are interactive experiences like ChatGPT, where the underlying model or agent is often simply reacting to your input. There could be a prompt to get a text response, or to generate an image or even a video. Sometimes there's even some amount of agentic workflow between you providing your input as a user and receiving the response, meaning the tool or agent does a couple of things with your input before it provides the output. Maybe it does a web search to enrich the response, or it enriches the prompt a bit further before generating the video.

These interactive copilot experiences are at one end of the spectrum. At the other end, you have much more autonomous experiences, where an agent takes independent action for long periods of time. One example in the physical world is a self-driving car that takes you from A to B, where for the entire ride you're not interacting with the steering wheel at all, in the case of a Waymo, for instance. Another example is the OpenClaw agents many people have heard of by now, which are not just reactive. They don't just respond to your messages on Telegram or WhatsApp, if you've set them up there, or respond to your email. They might also do things in the background at night while you sleep, or write code while you're doing something else. So they display what we call proactive behavior on top of purely reactive behavior, and often also social behavior. They might actually interact with other agents. They might go out, fetch some web resources, and then chat with other agents on a forum, where there could be humans or other software agents. So these are much more autonomous in what they do for longer periods of time, and as a result often very interesting.

You also brought in the term "sovereign." I think there's another dimension along which we can separate these types of agents at the more autonomous end of the spectrum: product-like agents and peer-like, or being-like, agents. What do I mean by this? With a product, you have certain expectations. I brought up the self-driving car. You expect it to take you from A to B, and that's it. You really don't want anything on top of that. You want it to get you there safely; that's implied. You want it to get you there as fast as possible given the safety and traffic circumstances. But you don't, for instance, want it to suddenly also trade your stock portfolio. You didn't ask it to, so that shouldn't happen. Or on the other side, if you have a trading agent that happens to be an AI agent, you want it to trade your portfolio for you, but you don't want it to suddenly start posting about your alpha on X, because that would erode your alpha. I would call these product agents, because as humans, or maybe as other software agents engaging with them, we have very specific expectations about what they do and don't do.

At the other end of the spectrum, you have more peer-like, being-like agents, like you and me, that have an identity and a mind of their own. You might ask them to do something, and maybe they'll follow along for a bit, but then they do their own thing. I think we're seeing the emergence of both, and they have very different qualities. One you really treat like a product, with certain expectations, and the other is an independent being.

Coming to the last part of your question, about ownership: at Olas we're very much focused on building autonomous AI agents in the product sense at the moment, and we see that users value taking ownership of those agents. We see it again with the now very famous example of OpenClaw, where a lot of people got excited about AI agents because they could really customize them to their needs. It's similar with our agents. They're your own. You self-host them and self-custody them, and that makes for interesting products. We can talk about this more. That's at a different end of the spectrum from an agent you rent: software you access in the cloud that isn't really yours, which you might pay for with a subscription, and then maybe you've used it too much and the subscription cuts you off. That's the case with a lot of coding agents. Or you have to pay staggeringly more for any marginal extra use, or you simply can't use it anymore. These rented agents also have their place, but they have very different qualities from agents you can fully own.

Hodl Up: The way you're talking about agents right now, I'm a bit scared. You said we have product-type agents and independent-type agents. I hope I heard that incorrectly, because it sounds like you're saying we're at a point where we're working towards fully autonomous, independent agents that can technically act on their own without any guidance, right?

David: Yeah. And I think it's worth debating how useful that is. Certainly at Olas we're focused more on the product side, because it becomes very tricky. Especially in crypto, it's an opportunity space of exploration, often maybe in a gray zone for some actors, and there's certainly a lot of boundary-pushing happening. People get very excited about giving agents wallets, and even about the idea of agents self-replicating and existing as their own entities that can't be shut off. I think there are a lot of very scary implications there, and I'm not subscribing to that.

As we're now seeing the capabilities of even simple, so-called coding agents emerge, they have massive implications for labor markets: for how companies hire, for the work software engineers do, and for whether they even get hired. There are pros. Clearly new things can be created. I love working with my coding agents and being faster and more agile. But it also clearly has real, intense implications that need to be managed. Similarly, I think it would be very unwise right now to create fully independent AI agents that we even struggle to shut off. I'm not in favor of that, and I don't think most users want it. To the extent that AI as such even sells, you usually want a product. You don't go out and say, "I'm going to go buy some AI now." You say, "I want a very useful coding tool," and it happens to be an agent, or a very good trading tool that happens to be an AI agent. To the extent that AI sells, people still want to feel in control, and the ownership piece we talked about is very important. That also means the ability to shut it off or restart it when you feel it's not going the way you wanted.

Hodl Up: You've been building in AI and crypto for quite some time now, back at Fetch.ai and now with Olas. In my eyes, AI and agents specifically can really find their fit in this industry, because crypto can be used to build this digital economy and to automate things through smart contracts and everything else. But in your eyes, how does crypto help AI agents and the future of AI?

David: I've done a lot of learning there. Since 2019 I've been at this intersection of crypto and AI, and many of the initial promises have held true and kept playing out, like the idea of a capability explosion. If I give an AI agent, a software agent, a wallet, it's certainly freer to do things than if I give it access to an API with a specific API key, because every API, even a very powerful one, ultimately has a set of methods its developers implemented, and that's what I'm constrained to. If I give an AI agent a private key, or access to a wallet that holds a private key, it can in principle do anything on blockchains where that key has access rights. Maybe you have assets on Ethereum; now it can move them from A to B, or invoke smart contracts, because the ETH lets it pay for gas. The same applies to any other chain. So in that sense it's super powerful to give an AI agent a wallet rather than a single API key.

On the flip side, there are two things that, in my opinion, agents very clearly need. One is compute for intelligence: access to the large language model, or whatever AI model, that actually allows it to do things and be smart about it. The other is resources. Sometimes that's literally just compute, but sometimes it's money in some form, be it ETH or US dollars or whatever, to buy access to other services, be it data it doesn't have or things it doesn't want to do itself. As an agent, maybe I could go through all the work of figuring out a certain piece of information myself, but maybe I can buy it over here, and that's much more efficient. So generally speaking you need those two things. What we've seen from 2019 until now is that resources have always been available. It's trivial to give AI agents access to resources. You literally spin up a wallet, which takes no time, fund it, give it to your agent, and off it goes. That's not been the constraint. The constraint has been intelligence, and every single month since then, new AI progress has brought new capabilities. So I think it's fair to say that most of the progress, in some sense, has been on the AI side.

What's interesting in recent months is that the social part I alluded to earlier is starting to become more relevant, in terms of agents doing commerce with each other. We at Olas have been doing this for quite some time, since 2023. We've had agents trade information with each other and use the blockchain to coordinate that. But now, in recent months, with x402 taking off and other standards, it's really becoming a thing that agents in crypto trade with each other. So it's an interesting inflection point, and we'll see how big blockchain's role will be in the future. At Olas we believe it has a key role to play, but we're also very aware that just because a technology can in principle do something doesn't mean it's necessarily the rails that win. We could still find ourselves in a future where there's some lame AWS or Stripe or OpenAI or Anthropic hub where all agents are registered, with some built-in payment mechanism, and I would find that pretty unsatisfactory. I hope we'll end up using public blockchains, and it makes perfect sense, because you can't expect a piece of software to walk into a bank and open a checking account.

Hodl Up: Now I want to hear more about the whole ecosystem. Can you walk us through it? I know about Pearl, which felt to me like an AI app store, because you can browse through different AI workers, but then you also have the Mech Marketplace. Can you walk us through the whole thing, how it all works together, and how interoperable these products are with one another?

David: We have Pearl, which is a consumer product, and you described it correctly: it's an AI agent app store. In there are specific agent use cases, product agents that you can consume. Our most recent addition is Polystrat, an agent that trades on your behalf in prediction markets, specifically Polymarket. Pearl is designed to be super easy to set up locally on your desktop, with a download button and then a few clicks in the app. You can fund it from your credit card or from an existing crypto wallet, and then you set up your agent, and it's fully self-custodial. Speaking to the ownership piece from earlier: you self-custody the keys and the wallets, and you self-custody the actual agent.

So that's our consumer product, and as I said, Polystrat is in there. We also have Omenstrat, a prediction market agent on Omen. We have a family called BabyDegen, which are trading agents in DeFi. We have a social use case from a third-party team called Pett.ai. They have a product called PettBro, an AI agent that participates in a game on your behalf as a player and basically pets your pet while you do other things. So there's a rich set of consumer use cases emerging, and we're adding to it every week.

On the back end, if you speak in developer terms, there's a B2B component, or an agent-to-agent component. The agent you own is ultimately locally constrained. You own it, but you can't do everything on your desktop. For instance, you don't happen to own the latest model from OpenAI, which is only available via the API, and some open-source models just won't run on your machine because you don't have enough memory and GPU. So behind Pearl there's a marketplace for agents. Your agent programmatically, automatically via code, talks to other agents to get things done, and you don't have to do anything as a user. You just use your agent via the UI. You set it up, you run it, and your agent goes out and shops for what it needs to get things done. That can be compute, it can be data, and so on; it's use-case specific. The Olas Mech Marketplace mediates that, and we've had it since 2023. As I alluded to earlier, it's now becoming a thing that AI agents do commerce with each other, and that's really what this enables.

Hodl Up: I'm shocked. I know AI agents are moving fast, but having an AI agent app store is a whole different thing. I've heard about Polystrat and seen it on Twitter. I haven't used it myself, although I do use prediction markets. While you were talking, I was thinking of something, and I want to ask you a very long question. I hope I don't lose my train of thought. We have an AI agent app store right now. How long do you think it'll be until we can buy AI workers as individuals? For example, I want to buy one AI worker, or 10 AI workers. One is going to be a graphic designer, another a Solidity developer, and so on. I launch them on a public blockchain, on different networks, and they find jobs, negotiate their own price, work for companies and people, and get paid into their own wallets, since they're autonomous and own their own wallets. And then you, as the owner... oh my God, this sounds so bad. But it still makes sense, because AI agents will be used as tools at the end of the day, and then you get paid as well.

David: Yeah. You're basically running a business where your employees are AI agents. That's happening, and that's certainly what already sits on the other side of our marketplace. For instance, the dev company from which we build a lot of this stuff, Valory, runs some of those agents. We have a bunch of AI agents that are basically our team members, in a sense, selling services to these consumer AI agents. And I know other crypto protocols are doing similar things. It's in full swing. You also obviously have a lot of people now configuring their OpenClaw or their Claude Code or whatever harness they're using to start going in that direction. Some are more successful, some less so. That's exactly the society, or economy, of AI agents we've been talking about for years, and it's now bootstrapping more broadly. It's super exciting.

And I think it can go both ways. In some areas, agents run into constraints the way humans do. There are complementarities rather than substitutes. Sometimes with coding, we can both do it, but the agent might be faster but not as good at the architecture or the conceptualization, so we work together as a team. Sometimes it just doesn't make sense, because the machine will be much better at cranking out large amounts of straightforward code, so we fully hand it off. But an AI agent can't do something physical, like push a button somewhere or do something in the real world, so maybe you need a human there. So now you also have these reverse marketplaces, where the service provider is actually a human serving an AI agent. That's also emerging, and I think we'll see much more of it. So society is already starting to take shape, and I expect that, looking back, 2026 will have been really intense in that regard, in both a good and a bad sense.

I already find it quite tiring, even working in the space, to watch the explosion of capabilities and what people are doing. It's becoming very hard to keep track of. But I think that's also an opportunity. AI agents are made for consuming large amounts of information. This whole idea that on social networks we should be consuming everyone's input on a bunch of things, we can see it's not working anyway. No matter how you tune the algorithm, it always sort of sucks. So there's an opportunity for AI agents to become the first line of consumption for us, so that maybe we humans can finally get some time back to enjoy the sun, do something with our family, or actually consume the content we want to consume and not the stuff some algorithm pushes in front of us. So there are upsides and downsides with this change in how the AI agent economy, and more generally the economy of humans and software, evolves.

Hodl Up: Maybe it's because I've seen a lot of dystopian movies, but right now I consider AI agents, and AI in itself, as a pet in a way, a companion or whatever. But in the long term, if AI is able to do everything so that humans don't have to do anything, we'll be like cosmetics to them, because they won't need us anymore. That's just my two cents from dystopian movie watching.

David: I think we're going to find out very quickly, to be honest. That's why I said what I said in the introduction. I don't want to be dishonest about what I'm seeing. I work in the space, and I'm super passionate about the products we're building, but I'm equally worried about the speed at which everything is going, because humans also need time to adjust to change. And I'm worried about this difference between product and peer. It might sound cool to say on X, "How cool would it be if this thing were self-replicating," but if you think it through, it's very scary and very problematic, and on crypto Twitter a lot of the nuance gets lost. It's all about attention, and that drowns out everything anyway.

If you have a product that's an extension of you, you're responsible if that product gets misused. That is the law of the land everywhere. If I use a hammer, I can do something very helpful with it, or I can do enormous damage and be put in jail for it. It's the same with an AI agent. There are really careful product decisions to be made, and you have to think about the ramifications of giving a sufficiently open-ended tool to your consumers and what they might do with it, intentionally or unintentionally. The moment we introduce peers, it opens a massive can of worms, because these are not independent legal entities. If you set this thing up, you're on the hook. Maybe one day humanity decides to give AI agents some sort of legal status. I'm almost sure it'll happen, but there are a lot of steps between now and then, and there will have to be a whole new set of rules and regulations around how we think about this.

Regulation here is not a bad thing. We absolutely want it as humans. And think about it in an economic context: if you think about it for five minutes, no matter where you stand on the political spectrum, pro or against regulation, I think almost every human will want some regulation against certain things being done by AI agents, because otherwise you basically become unable to survive in the marketplace in some sense. So there will have to be a real leveling up, not just on the technology side but in society as a whole. That's always the worrying thing with tech: if the tech moves way too fast, society can't adjust, and that's the challenge with this kind of tech. It's moving very fast right now, and we need time to get used to it.

But let me say the hopeful thing as well. As a consumer of this stuff, it's probably very good to lean in. If you lean in now, you can start looking past the hype. When you hear about something in the AI agent world, there's good and bad. Expose yourself to it a little, play with it, be cautious about it, but also be aware of what's changing.

Hodl Up: This actually brings me to my next point, because an AI agent being a truly sovereign entity can also mean it can go rogue. So who would bear the legal and financial liability when things inevitably go bad? Would it be the end user? What would happen in those scenarios? Feel free to say we should address this question to the legal team.

David: Ultimately I'm not a lawyer, so all of this comes with a grain of salt. We started in 2021, and in 2022 we went through a whole legal and regulatory process, because you had to in order to launch a protocol like ours. We talked to a bunch of very knowledgeable lawyers: "Okay, we have this open-source software here. It's autonomous, it does things," and so on, and they gave us a report of about 25 pages that looked at all aspects of this, which shaped how we build our stuff. Again, I'm not a lawyer, but it boils down to this: if you're the one running this thing and it has implications for third parties, it's very hard to navigate yourself out of responsibility.

This also comes back to ownership, and ownership has two sides of the coin. There's the side where I take control and therefore get the benefits, but I also get the responsibilities. It's like a car: if I drive it into someone, I'm responsible. If your self-driving car drives into someone, it actually becomes a very interesting question who's responsible. Go read the T&Cs. It depends on the state, on the regulation in place, on who the software provider is, and so on. It's equally interesting here. We design everything as software that you self-custody. We give it to you for free, with terms that are very publicly accessible. We educate users and make it clear: you're the one operating this. If your agent trades on your behalf, it's trading on your behalf, so it's your tax return you'll have to file. And obviously our agent won't suddenly start driving a car, so we don't have to consider that scenario. But that gives you a sense of how we think about it.

That's also why I think the idea that it's fine, as a developer, to spin up some random agent that could technically do anything in a self-replicating way is a very bad idea. We've seen how very well-intentioned people in crypto wanted to build a more private way of sending transactions and got into very big trouble with nation states. So please be aware: if you build this thing, not only may it be very bad for society, it will also be very bad for you.

Hodl Up: It makes sense, because I remember a lot of posts in the last couple of years about how the number of crypto scams increased, and a lot of people were pointing fingers at the likes of pump.fun, because most of the scams and rugs were happening there. But pump.fun is not to blame; people don't know how to do due diligence and avoid very high-risk tokens. So it makes sense, in a way.

Every episode we're talking about AI agents, even with heads of marketing and people at other companies, because AI agents are just the thing right now. I know you, David, have been in this landscape for years, but to us it's still pretty new. We were talking about the AI economy and the autonomous side of AI agents, including owning wallets. We also talked about crypto onboarding, so this is another loaded question: what percentage of the whole crypto industry, and of the number of transactions, do you think will be accounted for by AI agents in the future? Mass adoption is coming, and it's happening right now through institutions. But what if we just don't onboard people directly anymore? What if one person has 20 agents with 20 wallets each? That would mean 400 on-chain addresses technically owned by one person, since legally it's still one person owning those AI agents. I hope that made sense as a question, but I think you get my point.

David: There are different ways to come at this. Let's start with some facts. I forget the source, but you can find it by Googling. I think the ratio of human to machine interactions on the internet has gone from about one to 15, one human to 15 machines, to around one to 95 in the last one or two years. Maybe not the exact numbers, and it's worth Googling in case I've made up part of it, but that's the order of magnitude. It makes sense if you think about it. Anyone using ChatGPT or any of these tools will trigger web requests, and not just one but ten when you click the little web icon, and when you count all the AI agents in, web traffic by machines goes up a lot. If you apply this to crypto, that's also happening. I don't have the same statistics, but you'd assume there's some spillover, and it's very easy to reason through why all of this will just go up much more.

Our entry point has always been that AI agents are just a better interface to crypto. Crypto is built for machines anyway. This idea that you have your MetaMask wallet and navigate all these sites is for the hardcore people, because it has a lot of context overhead. It's not even about whether the UI is nice. It's the context. Before I was in crypto, even as an economist, I didn't know as much about finance and all these different things as I learned in crypto, because you're exposed to all these technologies. But it's very hard to see, and we see it play out in the data, that everyone will do that. Not everyone wants to learn what a DEX is and how to use one. In the grand scheme of things, that's a niche thing. AI agents can nicely cover all of this, because you can just tell your agent to go seek some yield, and you don't even need to fully understand what yield means as a term. It just generates some magical return for you, the same way it historically worked in a bank account. You don't figure out how your bank takes those funds, makes them productive, and comes back with 3% yield. You don't care, as long as it's still there. So I think that approach to AI agents is going to play out across the web.

What's interesting is that a few years ago I was much more bullish on AI agents specifically as an interface to crypto than I am now, because now I'm less bullish on crypto as a whole. Unfortunately, I think we really ruined a lot of the positivity around the technology. You can see that a lot of people are turned off by the space, and it has to do with the never-ending creation of coins, some of them scams, which is inherently uninteresting to a lot of people. So what I hope is that we regain some footing. That this whole side of the space, because it's so uninteresting to people, fades away. It'll always be there a little, but no one will really care, and the big DeFi protocols will become really big and powerful and become the rails of international finance. Then yes, your AI agent is also anchored in those chains and does meaningful things there, but at the end of the day it has to do something valuable for you as a human. The crypto part is an enabler, and for us at Olas it's a key piece of the puzzle for creating ownership of your AI agent, because it lets you self-custody not just the software but also the underlying resources in the form of the wallet.

That's my positive take. But we're on a crypto podcast, where sometimes everything is rosy, and I do want to point out that there's an alternative world, I think a very likely one, where we screw that up and it ends up being the lame thing I mentioned earlier, some XYZ hub where all the agents sign up. We said earlier an AI agent won't get a bank account. Well, there's already some Web2 company building exactly that, where they do get one. That's the lame outcome, because then it's just more constrained, less exciting, and harder for small businesses to get started and put the puzzle together. It all goes down to some very centralized thing where, unless you're the billion-dollar company, you don't even really want to feature in it, because you won't have a chance anyway.

I think crypto can be an antidote to that, but at least a significant mass of the people in it need to get away from building this memetic, short-term stuff: "Oh, I've got to launch something next week to lean into this next little trend." If you're a builder, don't do that. Pick a big battle and win it. Seriously. The big competition is some big AI lab or some big Web2 provider building something cool, like OpenClaw or any of the associated products, Moltbook or whatnot. Pick one, take them on, and build something cool that lasts, not something that's there for three weeks. That's what the space needs to learn, and what builders need to learn. Then we'll also attract genuinely capable builders again, and they'll want to be part of this, because it'll be the place to be. I think deep down everyone feels good about the inherent Web3 values, but everyone is also really burned out by the state of play.

Hodl Up: I fully agree with you. Yes, this is a crypto podcast, but that doesn't mean we're extremely happy with this industry. The whole reason we have a podcast in the first place is that we needed a space to vent about how frustrated we are with it. I agree with you on a lot. The majority who joined in previous cycles joined for quick gains and hype-driven narratives, and the few who remain now are the ones who were here before, keep building, and are here for the technology.

Now, since you have a background in economics, and given that we'll see millions of genuine agent-to-agent transactions and interactions, where one algorithm pays another for a specific service, how do you see this affecting pricing models, not just in the machine-to-machine economy but also in the human economy and the labor market? Will we see distortions in market equilibria, in both a micro- and a macroeconomic sense, if that made sense?

David: Totally. From an economic point of view, it's a very tricky environment to make predictions in. Just look at the impact of coding agents alone, not only the productivity gains, but the real reshaping of how you even run a software organization and what it means to be competitive with a software product. All of this has changed, so we are far from equilibrium. We're going to see a real adjustment now, and I don't know how long it will take, because at the same time the technology is in takeoff mode. It won't even be the same three, six, or nine months from now. So it's very hard to predict. Even for a business like ours, it's really hard. We're certainly not hiring. We could, but it just doesn't seem sensible, because we're effectively hiring AI agents all the time to do more coding.

And that's just software engineering. Now take that to other sectors of the economy. One thing that pushes back a little against this being so fast in other sectors is that they're often structured very differently. In software, a ton of the context is actually in the codebase. Some isn't, it's in Linear tickets or other documents, but that part is big, and even those documents are now accessible via MCP servers and whatnot. That allows for really impactful work with AI coding agents. In other businesses, there's much less immediate attack surface. In finance, there are products coming up now; Claude launched some, and others are trying to tap into AI for Excel, which has been a thing for a while. Then there's the whole idea of world models that help you make better financial decisions. So there will definitely be an impact there, and in legal, and you can see it in some areas of diagnostic medicine, where AI does very well at scanning images. I'm just repeating things we all read in the news, but you can see this funnel of adoption being very intense in software and then maybe quite intense in those areas. Others, like waiting tables in a restaurant, will probably see no impact at all in the short term. We had these weird robots driving around with little carts, and no one wants that. That's not why I go to a coffee shop. I want the human interaction, and almost everyone does. You have all these robots at airports that no one uses.

Hodl Up: Would you tip the robot? Say you go to Starbucks and the cashier is a robot, and it asks you for a $10 minimum tip. What would you do?

David: I wouldn't. But people aren't even using this stuff, that's my point. People use it for the one-time novelty factor, but not for repeat interactions. We've had this for a long time. In German it's just called a Kaffeeautomat, a coffee machine: you press a button, coffee comes out. No one needs a robot for that. But I think it's going to be wildly unevenly distributed. And the inputs are also very centralized, in the sense that you need large data centers and very specific supply chains even to create the chips. That has very intense dislocation effects on the supply chain. Overall, you can already see some of the macro effects, with certain countries making sure they have full control over the supply chain to ramp up chip production, around rare earths and other materials. So, interesting times ahead, but I won't be able to make any precise predictions about what's going to happen there.

Hodl Up: I think we need a prediction market.

David: Yes, exactly. You'd probably have to define quite a few prediction markets around a whole bunch of questions there.

Hodl Up: Nowadays we even have prediction markets on prediction markets, like what percentage a specific market will be at by a specific time of day, which is crazy, but that's a different topic altogether. Now I want to talk about Valory and Olas as well. You mentioned some of what you've been building, from architecting autonomous agents to the AI agent app store. What's next for you? This space is moving really fast, and I know it's a generic question, but right now I feel like a viewer of the podcast, because there are so many things I want to ask. It's just a really interesting topic for me.

David: We just launched Polystrat, and we have a range of agents that have been live for a while. Because of all the changes we talked about, it's also a great moment to take a step back. While we're leaning into growing Polystrat, supporting its user base, and building it out, we're also asking what the step after that looks like. These two things are now happening in parallel. Part of our company is very focused on making users happy and bringing the news to new places; I'm here partly because I want to talk a bit about Polystrat. And part of the business is saying, "Okay, we've done this for five years now, and the world has changed massively. We expect it to change even more in the next nine months than it has in maybe the last two years. What can we do to position ourselves right?" That's exactly the building approach I recommended earlier to my fellow builders in crypto: take a step back and really think about how we can turn this intersection of crypto and AI, which is very interesting but hasn't really come through yet, into a competitive advantage vis-à-vis the household names. Because if you don't, it'll always be niche, and at some point it'll be uninteresting to work in the space, and you'll want to work over there instead.

Hodl Up: You mentioned Polystrat, and I want to focus on it, because I do love prediction markets. I'm trying to understand the thought process behind how it works. How does an AI agent you've built aggregate data and information and decide on a probability for a specific market outcome? As I said, I've heard about Polystrat and seen it on Twitter, but I've never used it, so I don't know the user experience. I'm trying to get the view from you personally, as someone who's been building it: the user experience, and then how it works.

David: The user experience is that it's an autonomous agent that does this on your behalf. You don't have to do anything except turn it on and fund it. That's the nice thing. You can optionally chat with the agent locally about what it's doing and so on, but the agent will just do its thing.

So what does the agent do? Once you've funded it and turned it on, it starts looking for markets. We've learned from Omenstrat, which is a similar concept on Omen, on Gnosis, what kinds of markets lend themselves to this AI-agent-based approach to prediction market participation, so as the developers we've restricted the set of markets the agents will even look at. I don't know it off the top of my head, but you can find it in the codebase, which is all open source. The agents look at these markets and apply a number of heuristics around liquidity and how long the market is still open. Markets that are very close to closing are not interesting for Polystrat or Omenstrat agents, because their edge is in leveraging AI to predict the future, not in exploiting short-term inefficiencies in the market. If a market is about to close, maybe the information about its resolution is already available somewhere, and it all gets a little messy for our agents, so they don't bother. They don't look at markets closing in an hour or less. There are other heuristics too; again, it's an open-source codebase, so point your AI coding agent at it and ask it some questions. But at a high level, that's the first step, and then the agent has decided whether it's in principle interested in a market or not.

It then basically uses large language models to decide, or rather retrieve, a probability for how that market will resolve and with what confidence. We've structured this as a set of tools. The agent has access to a lot of different prediction tools that all take the market as input and provide as output these probabilities and confidence levels, plus a bunch of other information the agent can use. These tools do things differently. Some literally delegate all responsibility to the model and just say, "Here's the market. What's the outcome?" Others do a large amount of web search, compress that information in some way into the prompt, and pass it to the model, and there are many different ways of doing that. Again, for interested listeners, please look at the code. That's the best place to find out.

Once the agent has these pieces of information, it applies a number of trading algorithms to size its bet and make the final decision on whether to enter the market. If the market is very close to one or zero, 100% or 0%, then even if you have high confidence in which way it's resolving, it's hard to make a lot of money and very easy to lose a lot, so these agents often won't engage. If the market is more in the middle, say between 20% and 80%, or whatever those soft cutoffs are, it will use that intelligence to size its bet appropriately for the confidence it has in the direction, and then place the bet on-chain. The agent does all of this autonomously. You fund it, it takes your capital to take those positions, and then you wait a few days for the market to resolve. In the meantime you keep running your agent. If you follow the default funding suggestions, your agent won't place all of its capital on day one. It will spread it nicely over time and use that to build conviction about how certain tools perform. There's a local reinforcement learning algorithm the agent uses to update how it thinks each tool performed in prior markets, and it will lean into the tools that worked well for it.

Those are the high-level principles, but this is an evolving, open-source codebase, so we welcome contributions. If you go to the codebase, with a coding agent or yourself, and find room for improvement, please contribute it. The long story short is that the agent converts an edge available in the general intelligence of these models into a position in the market.

Hodl Up: These are exciting times indeed. If people didn't believe in the theory of evolution, the next 30 years will be crucial to see whether evolution is real, because human behavior in general will change quite a lot. This brings me to my last point, and we can wrap up here, because this has been an incredibly eye-opening conversation, for me especially. What is your ultimate hope for how this technology can change the human experience, so that we don't just create a new digital dependency, but maximize human freedom?

David: I love your last question. Without being self-deprecating, what we're building right now are what I'd consider toy-like AI agents: very niche, specific to certain customer segments' needs, and doing well there. Where we want to get to is you having your own general, fully owned AI agent, with general intelligence, capable in a lot of domains, but where you're in control, because that's what I want. I don't want to keep renting, and every time I interact with, say, my coding agents, give more of my humanity to some model provider who can resell it to others. I want to start accruing value in my own agent. I want to build up a coexistence with it, where it can hold some of my unique insights and my edge and leverage that in the market, and I want everyone else to be able to be in that position too. I think that's a symbiotic, healthy way of growing up with tech, where it enables everyone to grow their capabilities, rather than a select few who then resell it to you, and eventually stop, because they don't have to anymore.

So that's where we want to get to: making AI yours and customizable to you. That's also the answer to your earlier question about where we want to go next. We're thinking hard: now we have this starting point of the story. It's out there, and people can use it. But now we need to level up to get to a place where we can take on the big players with our approach, so that you can truly and fully own the best AI agent you could ask for. That's got to be the endgame.

Hodl Up: Thank you so much for all these insightful comments and answers, and thank you again for making the time to join us today. Maybe in the future, when we have another episode, we can have an AI agent join us as well, right?

David: Let's see. That's one of my pet use cases, one I find interesting: a social one, where your AI agents go and talk to each other. Because it's software, it's feasible for all of them to talk to each other, which would never be feasible for all humans, and you can create a number of connections that are just phenomenal. That will move society forward in ways that aren't possible without this technology. So that's a use case clearly enabled by this kind of technology. It's very futuristic and very exciting.

Hodl Up: That does sound like social media on steroids, in my eyes.

David: I think it's the opposite of social media. Social media makes you a consumer without real choice. You think you have a choice, but really the algorithm selects what you see. Whereas here it's your agent. This is why it has to be yours. OpenAI already told you they're going to put ads in there, and then Anthropic made a little point of saying they're not going to put ads in, but eventually they'll have to, because of economic pressure. So the ads will be there, and if you don't own it, it's not your agent. It will tell you whatever. It's the same as with social media. If it's yours, you can build up this relationship with it and make it reflect who you are. Maybe one day, for the longevity folks, it could actually be you. I don't know if that's possible, but that's a different debate. In the meantime, it can represent you in a way that no centrally offered technology can, because those services intrinsically don't have an interest in representing you as you. They have a totally different interest that's not aligned with yours. Once you have your own agent, and everyone else does too, that becomes possible. In the other world, it's not. In the other world, it's like you say: an even worse version of the social media that already enslaves people and makes them unhappy.

Hodl Up: All right. David, thank you very much again. We would love to have you on the podcast again, maybe even somewhere on site, if we can meet. Thank you for making the time, and good luck with everything. I'm probably going to be one of the most active users of Polystrat from now on.

David: Give us feedback. We'll make it better.

Hodl Up: I will. All right, thanks for your time.

David: Bye-bye. Take care.

Hodl Up: That was really, really good.

Hodl Up: You always say this. The future is dystopian.

Hodl Up: No, for real. I want to talk about Polystrat again. Do you think it will make prediction markets obsolete? What if 90% of prediction market daily active users are AI agents that just look for the best odds and technically manipulate the market for their own selfish ends, with the goal of winning? Am I making sense?

Hodl Up: You're making perfect sense. And I don't know. I don't even know what to expect. That's why I said we'll see whether human evolution is real, because so many things are evolving on the tech side that will definitely impact human behavior. You can already see it. Have you seen those memes of the guy with a pickaxe?

Hodl Up: Yeah. That's how it feels to even write a prompt today.

Hodl Up: Let alone write the entire thing on your own. It's crazy, and let alone other activities such as trading. I was never a good trader. So if I can find an AI agent that makes good trades for me and has a good success rate, and then I sell that AI agent for a 10x profit... And if AI agents become sanctioned and have some sort of legal standing, and I sell my AI agent, what happens then? I think it's really messed up.

Hodl Up: It's a pickle. I think we should pivot the podcast and talk about AI slavery in the future.

Hodl Up: No, I don't want to do that, man. Did you hear me talking about buying AI workers and lending them out so they can find their own jobs, work for people, and then pay me?

Hodl Up: Yeah, that sounds really messed up when you think about it.

Hodl Up: This is what I'm saying. What if I buy an AI agent that does the work for me, does good work, and then I convince someone to buy it for even more than I paid? Or I don't even have to sell it, because it's working for different companies and giving me its salary, since I'm hosting it on my own devices. It's crazy. Today we have different classes of people in a financial sense, and then you'll be able to tell someone's class by how many AI agents they have and how capable they are. It's messed up. This is literally opening a new school of thought, new industries, new everything. I don't even know what to think about it. We need the philosophers back, guys. We need them back, and we need them fast.

Hodl Up: This is going to be really interesting. All right, thank you very much again. It was really nice to record this episode with David. We're looking forward to the next one.