Pearl, prediction markets and the long tail of AI liquidity

Pearl is Olas’s consumer gateway to a future where narrow AI agents quietly trade, curate and create prediction markets at a scale humans will never touch, says co‑founder David Minasch.

David Minasch sat down with crypto.news on March 31 on the sidelines of ETHCC to explain why Pearl’s narrow, long‑running AI agents are remaking prediction markets from the inside out.

From Fetch.ai to Pearl

Minasch’s route into autonomous agents is textbook crypto‑AI convergence. “I got drawn into the space by my background in economics and game theory,” he told crypto.news, recalling his move into crypto after several years working on machine learning applications.

At Fetch.ai, where he spent two years, his team “built the first agent framework in crypto probably,” anchored on a simple but loaded idea: “you’d have sort of wallets that not be controlled by humans but by machines.”

“We actually wrote a detailed paper on this, which was way ahead of its time,” he adds. In 2021, he spun those lessons out into Valory, the core lab behind Olas, which has since experimented with a range of applications and go‑to‑market strategies.

The first bet was B2B: autonomous agents sold to DAOs such as CowSwap, Balancer and Ceramic. “That went okay but never sort of really took off,” Minasch concedes. The real pivot came in 2023, when “general purpose usable large language models like ChatGPT” landed and Olas “switched more to B2C.” Pearl is the result: “a B2C application which has different agents in it,” built for users, not governance forums.

By the time Pearl launched in February 2025, the rest of the industry had caught up to Olas’s early agent thesis. “The crypto space and the AI space had moved towards agents, now everyone is building agents or using agents or both,” Minasch says. But he argues most people’s idea of an agent is still shaped by chat interfaces like ChatGPT: “a co‑pilot synchronous experience” where you prompt and it replies, in front of you, in real time.

Olas is explicitly betting against that dominant pattern. “When you have long long‑running agents with like autonomy but tightly scoped so they can’t just do anything but they can do interesting things within a certain scope. That’s where it becomes very interesting,” he says. Pearl is designed around those tightly scoped, background processes rather than generalist assistants, Minasch points out.

“With Pearl we intentionally go very narrow in terms of the capabilities of an agent,” he explains. He points to new tools like OpenClaw—as both validation and warning. “OpenClaw validated a lot of our core assumptions that people do want llocal first experiences with AI agents,” he says, but “the product can do too much, which causes a bunch of problems, including secruity, but also just a problem for the user.”

In his view, that kind of system is built for tinkerers “who just sort of want to mold this thing into something that’s useful to them.” The “low friction user” wants to “just press a button” and get a consistent result. “I have one and I asked it to send me daily report and half the time it’s broken,” he says of OpenClaw. “That’s not a good product experience.” Pearl’s agents, by contrast, are designed to do one thing—trading, yield seeking, market creation—reliably. Limited scope, high definition, low problem latency.

Polystrat is the cleanest demonstration of that philosophy. Polystrat is an example because here’s just the idea: provide some capital, have it trade in prediction markets,” Minasch says. Instead of facing Polymarket’s UX—wallet setup, funding, market selection, position sizing—the user delegates funds to Polystrat and lets the agent do the work.

“Polystrat is just like a user of Polymarket,” he stresses. “If you want to use Polymarket you as a human need to set up a wallet, fund it and then you’re faced with the decision of what market to trade in. Polystrat abstracts all this and the idea is for it to simply trade on your behalf.” The agent focuses on geopolitical and political news markets, “not so short‑lived” and generally closing “within the next four to five days.”

Technically, the flow is simple but ruthless. The agent filters markets using rules like liquidity and time to close, then applies “prediction tools,” which Minasch describes as “workflows that sit on top of models and data sources.” “There’s many different prediction tools and the agent learns over time which ones to take and which ones not to take,” depending on the market. A local pricing and sizing engine converts those predictions into positions and the system trades autonomously on your behalf.

Performance wise, Polystrat ranges between 56 and 69% accuracy, Minasch says. As a fleet, “our agents… have performed two to three times as well as human traders,” although they are “not yet at a fleet‑wide break even.” Individual Polystrat instances, however, can deliver “up to 100% ROI overall and like several 100% ROI per individual trade.” The goal is not anecdotes but a statistical edge: “to have a Polystrat fleet on average a positive ROI.”

Trading is only half the story. As more agents enter Polymarket and its predecessors, Minasch sees prediction markets becoming “early prototypes for these market‑driven AI systems… environments that encode truth discovery at an economic scale.”

He doesn’t pretend the rails are clean. On controversial questions—or markets with contested outcomes—information lags and disputed outcomes are common. Polystrat nor other agents on Pearl attempt to solve that. “Polystrat itself is just a trading agent on top of Polymarket,” it’s neither consensus building nor a truth serum.

But AI is already reshaping participation, creation and policing. “It’s unclear exactly how many traders in prediction markets are already AI agents but it’s probably more than 30%,” Minasch believes. “Potentially already more than half,” he adds. As such, humans have limited attention, so “the whole long tail of prediction markets will basically be served to AI agents,” he predicts.

Crucially, Minasch breaks from crypto libertarianism on governance. “We take the view that there should be regulation of prediction markets,” he says flatly, citing markets that “effectively look like assassination markets” or “incentivizing bad behaviors.” With “a certain degree of regulation or self‑regulation,” more markets and more AI participants should “drive prices to equilibrium” and “improve the information embedded in the markets,” opening the door to derivatives, hedging and other instruments built on top.

Asked whether Olas agents could become “data liquidity providers operating autonomously across multiple networks,” Minasch shrugs off the distinction. “Liquidity provision is effectively also trading strategy,” he says.

In that framing, Pearl is less a single app and more an operating system for narrow, long‑running agents: Polystrat for prediction markets, Optimus for yield, Omenstrat for market creation and whatever comes next for liquidity across venues. The consistent design choice is scope: each agent does one thing, over long horizons, with as little human intervention as possible.

“We were just very early to something that a lot of people are now doing,” Minasch says of the agent wave. The difference now is that Pearl is pushing those agents into retail‑facing products, turning prediction markets into both a playground and a proving ground for AI‑driven liquidity and truth discovery.