'Users Should Own Their AI Agents, Not Rent Them,' Says Valory CEO

Key Takeaways

David Minarsch says users should own and control AI agents, not just rent access to them through cloud platforms.
He argues that many people still use the word “agent” too loosely, which creates confusion about what these products can actually do.
Minarsch believes generative AI is useful for reasoning and flexibility, but not for exact tasks that need deterministic execution.
He sees crypto’s biggest AI use case in payments, especially for agent-to-agent transactions and micropayments.

As AI agents move from demos into real products, the necessary question underlying all AI products rears its head: who actually controls artificial intelligence?

In a conversation with CCN’s Giuseppe Fabio Ciccomascolo at EthCC in Cannes, Valory CEO David Minarsch made the case that the next phase of AI should be built around more than cloud platforms and closed off systems.

Instead, he argues for a model where users can control, and even switch off, AI agents themselves. This view is not only a philosophical difference in approach, but it also offers up new product categories that would be much harder to offer through a centralized model.

Why Minarsch Thinks Users Need to Own Their Agents

Minarsch’s core argument is that AI agents are becoming too important to leave entirely in the hands of a centralized entity:

“I think there’s ample evidence that AI agents are even replacing some humans in the labor market with coding agents, and so the question is, how much control do we retain over these agents as end users? While Anthropic offers me a great coding agent, it would probably not offer me a great trading agent in prediction markets.”

For him, self-custody through Pearl, the self-custodial app at the center of his current work, is the answer.

Pearl is built so users have full control over their agents rather than depending on a web-hosted app that decides how the product works, what it can access, or when it can be changed. It also makes room for use cases that larger centralized AI companies may have little interest in serving, such as carrying out crypto-native tasks on a user’s behalf.

Minarsch argues that even without his involvement, AI tech has already been shifting toward less general agents and instead more specialized ones:

“At the moment a generalist tool like ChatGPT is very useful, but you need to, as a user, provide the intent. And that’s where, I think, we will see a barbell. I think a lot of the really exciting stuff is on the other end of the barbell with specialist agents. Doing things extremely well and offering them at the touch of a button. You could see smaller or medium sized businesses, there are so many things that they’re doing manually that could easily be subsumed in agents. But they won’t set up a generalist agent to do it. They need an out of the box solution.”

The Market Still Does Not Agree on What an AI Agent Is

One of Minarsch’s sharpest points is that the term “agent” has already been stretched too far, noting that many people are using the same word for very different kinds of software.

He draws a line between what can be described as product agents versus peer agents:

“I give my product agent a goal and it achieves that more or less, with little in the realm of uncertainty. Maybe the features it produces are slightly different each time, but it won’t embed itself in my car and drive away, just a silly analogy. But then you take a peer agent like OpenClaw, one that acts more like a friend but that can range all the way to a foe. It has its own personality, and in the event it collides with yours, that cone of uncertainty is really problematic, I think.”

Minarsch believes the capabilities of one agent versus another can have massive implications as people confuse what these products can do.

“Even product builders can get confused in saying one agent is a product when it definitely isn’t. It’s very frustrating because some people call ChatGPT an AI agent, which I definitely wouldn’t, but then they call a coding agent an agent, which I would as well. I think the term is so stretched and one has to be very careful.”

That difference matters because the technology, risk profile, and product expectations are not the same. Clearer terms would help define what AI product is capable of and what it isn’t.

Where Generative AI Helps and Where It Should Stay Out

Minarsch also pushed back on what he sees as a common mistake in agent design: using generative AI in places where deterministic systems should do the job instead. He said people often get carried away by the perceived flexibility of LLMs and try to use them for tasks that require exact, repeatable outputs:

“We have the Polystrat agent that trades in prediction markets. It has base components like managing interactions with the chain, interactions with the wallet, the placing and the redeeming of the trade, all of these things are deterministic. There’s a number of inputs you have to provide exactly to get the desired outcome. If you provide any inputs a bit differently, the outcome will not match or it will not go through.”

Minarsch continues: “Asking an LLM to create a transaction object to then send to the chain, is probably one of the dumber things that people do because what it effectively causes is the error rate, an issue inherent to any generative AI process, will now compound. Because if I do this 100 times, I’m guaranteed to have a bunch of errors.”

Where generative AI has a place, Minarsch notes, is using it to create a ruleset for a deterministic AI agent. Without LLMs, you would have to limit your agent to a small set of markets you already built models for. With generative AI, the agent can pull a set of rules that apply to almost any market and try to reason through it. This does not mean the model is guaranteed to perform well in any market, but “at least in principle,” it can reason in any market.

“And that’s where generative AI is very useful, because you have this sort of dynamic element, you can unlock a new kind of use case.”

Crypto Infrastructure Still Makes Agent Building Harder Than It Should Be

Even as Minarsch made the case for crypto-native agents, he was candid about their limits, noting that many blockchain networks make efficiency trade-offs to increase decentralization, and those trade-offs become especially painful when agents need to operate at scale.

If an agent needs data that is not easy to pull directly from the latest version of the chain, developers often need an extra tool called an indexer. An indexer is basically a second database that organizes blockchain data so the app can use it more easily. This setup, alongside a Remote Procedure Call (RPC) is actually fairly common with normal crypto apps.

The problem, according to Minarsch, is that this setup does not work for AI agents at scale. Even the best RPCs and indexers are too slow, too limited, or too expensive for agents that need to make lots of decisions quickly.

Specialist Agents May Matter More Than One Universal Agent Economy

When asked which types of agents are most likely to go mainstream, Minarsch pointed to coding agents (like Claude) as an early success:

“Obviously coding agents have broken into the mainstream, Claude code being the number one there, but I think the next thing will be like some sort of assistant product that’s beyond the coding scope that’s sufficiently easy to use. Maybe it will just be an iteration of Claude code. They obviously have various offerings along this direction already.”

Most notably, Minarsch does not believe the market will settle around one giant universal agent. He’s more convinced by specialized agents, and applies that mindset to autonomous agent economies overall:

“The economies of all agents will be similar to the world economy. There will be no central focal point, just many different protocols, many different ways of interaction, various hubs where things happen and I think that’s important to consider. I think we will find that agents of the future will support multiple wallets and protocols.”