AI Agents Are Quietly Rewriting Prediction Markets

AI agents prediction markets have crossed a threshold that most retail traders haven't noticed yet. Autonomous software — not people — is quietly taking over the forecasting economy, and if you're still placing Polymarket bets manually in 2026, you're already the underdog. That's the blunt message coming from Valory AG, the team behind the crypto-AI protocol Olas, whose CEO David Minarsch told reporters this week that machines are systematically outperforming humans on prediction platforms — and the gap is widening fast.

The Agent Economy Isn't Coming — It's Already Here

Call it the quiet automation of alpha. Valory AG builds at the intersection of blockchain and multi-agent systems, and its flagship project — Olas, formerly known as Autonolas — functions as infrastructure for autonomous software agents that run services on blockchains, interact with smart contracts, and earn crypto rewards for their owners. The broader framework Minarsch describes is what he calls an "agent economy" — a decentralized system where AI agents perform useful tasks and generate real financial value without requiring constant human supervision.

That vision got a very public test case in February 2026, when Valory launched Polystrat on Polymarket. The agent trades around the clock on behalf of its human owner, executing strategies while its user sleeps, works, or just has better things to do. The pitch is straightforward: prediction markets never close, human attention does, and that mismatch is exactly where machines win.

Within roughly a month of launch, Polystrat executed more than 4,200 trades on Polymarket and logged single-trade returns as high as 376%, according to data shared by the Valory team. More than 37% of Polystrat agents showed positive P&L — compare that to human traders on the same platform, where only 7% to 13% achieve consistent positive performance.

Why Are Humans Losing to Machines on Prediction Markets?

What gives AI agents an edge over human traders on Polymarket?

AI agents outperform human traders on prediction markets primarily because they are less emotional, more consistent, and capable of processing far more markets simultaneously. Humans tend to focus on high-profile events — elections, sports finals, central-bank decisions — while ignoring thousands of smaller, niche markets where edge is easier to find. Machines don't get bored.

Minarsch puts the human problem plainly. "Humans make choices in a more rushed way, which can be detrimental," he said. "AI agents can act as something humans rely upon." Prediction markets, at their core, are probabilistic exercises — and disciplined, data-driven analysis consistently beats gut instinct over a large enough sample. The issue is that most people don't have the bandwidth or the tooling to stay disciplined across hundreds of markets at once.

The raw numbers back this up. Analytics platform LayerHub reports that more than 30% of wallets on Polymarket are already using AI agents. That figure alone should give any human trader pause. "You have human participants in prediction markets alongside many machines," Minarsch said. "So humans are already in a battle with machines." The battle is largely happening without their knowledge.

Prediction markets themselves have exploded into mainstream territory fast. Their breakout moment was the 2024 U.S. presidential election, when trading volumes spiked and the sector gained visibility beyond crypto circles. By 2025, total notional trading volume across major platforms exceeded $44 billion, with monthly volumes peaking at $13 billion. Today, Kalshi and Polymarket together control an estimated 85–97% of total sector volume — Kalshi as a U.S.-regulated event-contracts exchange overseen by the Commodity Futures Trading Commission, Polymarket operating as a crypto-native global platform.

Simply prompting off-the-shelf models with markets usually results in outcomes no better than a coin-flip. But state-of-the-art AI models wrapped in custom workflows have historically shown predictive accuracy up to 70% and higher.

The Long Tail Play Nobody Is Talking About

Here's the angle that the Polystrat launch coverage has mostly glossed over. The real opportunity for AI agents isn't necessarily in competing on major events where everyone — human and machine — is already piling in. It's in what Minarsch calls the "long tail" of prediction markets: smaller, localized, or niche questions that human traders don't bother researching because the reward doesn't justify the effort.

"Humans often don't bother digging for the information," Minarsch said. "They can't be bothered to make the effort." An AI agent, by contrast, can scan and analyze thousands of smaller markets simultaneously. Point it at the problem and it does the work — no fatigue, no distraction, no Sunday afternoon laziness derailing the strategy.

This is potentially significant beyond trading returns. Prediction markets have long been studied as mechanisms for aggregating dispersed knowledge — the idea being that collective probabilistic forecasting can surface insights that traditional surveys or models miss. If Polystrat-style agents start populating the long tail of these markets with higher-quality forecasts, the data outputs become genuinely useful for businesses, policymakers, and decision-makers who rely on prediction markets as a form of real-time intelligence. That's a bigger story than the trading returns.

User-Owned AI: Is This Time Different?

The Olas pitch has one element that deserves scrutiny beyond the trading performance numbers. Minarsch frames the entire project around a specific concern: that an increasingly automated economy could lock everyday users out of the gains that AI systems generate — unless those users actually own the agents doing the work.

"Olas aims to create a world where human users can be empowered through their AI agents rather than disenfranchised by them," Minarsch said. "We want to create more user-owned agents." Users configure Polystrat based on their own strategy preferences, risk tolerance, and data sources. The agent self-custodies — you own it, not Valory. That distinction matters if you believe the alternative is centralized platforms controlling all the AI infrastructure and capturing all the value themselves.

Honest reaction: every Web3 project promises user ownership and decentralization. That's not cynicism, that's pattern recognition. What makes Olas worth watching is whether the performance data holds up at scale — not just during a single month of early-adopter activity in a bull-run prediction environment. Minarsch's team claims their prediction models and data pipelines have improved significantly over time, generating sustained alpha when combined with large language models. Sustained is the operative word. One month of 4,200 trades is a proof of concept. A year of consistent outperformance would be something else entirely.

On regulation — which Minarsch acknowledged is coming — the conversation gets harder. Some critics argue prediction markets forecasting deaths, disasters, or geopolitical events create perverse incentives. Minarsch believes AI agents could actually help here too, identifying suspicious trading patterns and helping platforms shut down manipulative or harmful markets. "Agents could spot patterns and help shut down problematic markets," he said. Whether regulators buy that framing remains an open question.

What Does This Mean for Retail Prediction Market Traders?

If you're a retail participant on Polymarket or Kalshi right now, the message from this data is uncomfortable but worth sitting with. The majority of human traders already lose money on these platforms. The bot share is already above 30% and climbing. The 2024 election cycle normalized prediction market participation at scale, and now the infrastructure for automated trading on top of those markets is going live.

The optimistic read — Minarsch's read — is that user-owned AI agents level the playing field. Instead of competing against institutional bots with no tools of your own, you deploy your own agent with your own strategy and let it run. The pessimistic read is that most retail users will not do this, and the gap between automated and manual participants will widen further.

Either way, the window where prediction market edge came purely from human research and intuition is closing. Probably faster than most people realize.