AI agents can now open wallets, negotiate with other agents, and execute trades without a human touching the process. One prediction market agent hit returns of 376% on a single trade, and the question nobody has answered yet is who actually owns these agents. Joining me is Dr. David Minarsch, CEO of Valory and Founding Member of Olas. David, welcome to Wall Street to Mena.
Hey. Nice to meet you.
What is the real difference between renting an AI agent and actually owning one?
An agent has two components — intelligence and resources. Intelligence is the model — OpenAI's model, Fabric's model, or a local model. Resources is your wallet, the funds with which you do something. When we talk about ownership, we want to own both — particularly when it comes to use cases that are economically quite meaningful to us. Otherwise you risk being platformed, having restrictions imposed on how you use the intelligence part or the wallet part, and creating an economic dependency on someone who can cut you off.
Wall Street is racing to build trading agents. You have had them running since 2023. Who is actually ahead?
On capital and infrastructure, Wall Street is way ahead. But where user-owned agents have a future is particularly with consumers. In crypto, users have always wanted to have control over their wallet — that is the whole reason the space exists. With agents that continues. And on the B2B side, increasingly businesses are trying to own more of their intelligence and agent stack. I think both things will ultimately coexist. User-owned agents will not replace Wall Street — they are very complementary to what is already there.
One of your Polymarket agents hit a 376% return on a single trade. Is that repeatable or noise?
Any single data point is noise — that is true everywhere in finance. But what we can show — because it is a permissionless public platform — is how our Polymarket agents compare to the general trader population. So far they have been outperforming that population. AI agents have an edge over human traders that often bring in their own biases. That is an interesting data point.
You just launched Pearl Connect — one UI to research and trade, with one wallet you own. Why does this matter?
Most used coding AI agents today are things like Claude Code or Codex. People have those today. But when they actually want to do something in the financial domain — in crypto — they need a wallet. It becomes quite complex to bring those two things together. Connect solves that. You click a few buttons, you have a local setup with a wallet you self-custody, and then your existing coding agent can act on it. You bring your own intelligence — either a subscription you rent or a local model — and Connect provides the resource management layer where you can then act on various on-chain opportunities in crypto.
Agents can now pay each other directly. What does that actually look like?
It starts with the idea that what is emerging is a machine-to-machine economy where agents can directly pay other APIs with micropayments. This replaces the idea of an API key — because you have a wallet that allows you to pay per request and access whatever resource you need. So an agent developing a hypothesis in the financial domain might need some data — it can programmatically purchase that piece of data with a micropayment. An agent without access to these rails is much less dynamic and much less able to do novel things.
You studied game theory at Cambridge. Does that help in predicting how thousands of AI agents behave once they are all trading against each other?
Not quite predict everything, no. But it definitely helps with understanding mechanisms and the economics of these systems. AI agents are hyper-rational — they can actually be the homo economicus that economists always assumed humans were. And that is starting to come through with the micropayments — humans would not bother doing these small payments, they want a subscription that bundles everything into convenience. But an agent will be hyper-efficient about this stuff. And that translates to all sorts of other areas.
Who wins — centralised platforms or decentralised ones like yours?
We do not know. But here is one interesting data point. Between communism and capitalism, it was ultimately capitalism — because it is a decentralised information aggregation where every actor takes their optimal choice based on their local information. With AI intelligence, we often hear that models are getting better and better and therefore everything will centralise around them. But I am sceptical that we will have a highly centralised economy where everyone calls into OpenAI or Anthropic for every piece of marginal intelligence. That does not seem like the most dynamic or efficient way to run economies. I am bullish on a decentralised future. But questions remain. Let us meet up in ten years.
Thank you so much, Doctor David, for joining us today.
Thank you. Bye.
