Prediction-Market Agents in 2026: The End of Easy Money

The question about prediction-market agents has changed. Eighteen months ago, the interesting problem was whether an autonomous program could hold its own wallet, read a market, and place a bet on Polymarket without a human touching the keyboard. That problem is solved. By early 2026, more than 30% of the wallets active on Polymarket were running on AI agents rather than people, according to LayerHub data reported by CoinDesk. The machines are here, in force, and they are no longer a novelty. The newer question is sharper and much less flattering: now that the bots have taken the order book, can any of them still make money?

The evidence arriving through 2026 is contradictory in a way that tells its own story. The company behind the most widely used trading agent says its bots beat human traders by a wide margin. An independent academic benchmark that handed six frontier models real cash and let them trade for two months found that they mostly lost it. A study of 30 billion order-book events concluded that the public data feed most agents trade on is so noisy it gets the direction of a trade wrong roughly two times in five. And in August 2026, for the first time in a year, combined volume across Kalshi and Polymarket actually fell.

What follows is a look under the hood: the microstructure of the two biggest venues, the stack the bots are built on, the proprietary trading firms now crowding in, and the uncomfortable arithmetic of where an agent’s edge actually comes from. It is a companion to our earlier piece on how these markets decide who won; here the subject is the trading, not the judging.

From Novelty to Majority: How the Bots Took the Order Book

The clearest marker of the shift is Polystrat, an autonomous agent launched in February 2026 by Valory, the team behind the Olas network (formerly Autonolas). Valory chief executive David Minarsch describes it plainly: “In a nutshell, Polystrat is an autonomous AI agent that trades on Polymarket 24/7 on behalf of its human user.” Within roughly a month of launch, the agent had executed more than 4,200 trades and recorded single-trade returns as high as 376%, CoinDesk reported. It is one product among many, but it is representative: the barrier to putting a strategy-driven bot on a prediction market has collapsed.

The consequence is a market where people are already outnumbered in the places that matter. Minarsch frames it as a fight that has already begun: “You have human participants in prediction markets alongside many machines. So humans are already in a battle with machines.” That is not simply promotion. When a third of the wallets on the largest on-chain venue are automated, the counterparty on the other side of your trade is, more often than not, code that never sleeps, never tilts after a loss, and reprices the moment a headline crosses.

Agents are also drawn to a specific part of the market. “The long tail of prediction markets is very interesting for AI agents,” Minarsch has said. “You just point the agent at the problem and it does the work.” Humans cluster around a handful of high-profile questions, elections, rate decisions, a World Cup final. Machines can price the hundreds of thin, obscure markets nobody has time to watch. That is where the first wave of easy money lived, and, as we will see, where the professionals went hunting next.

Do Prediction-Market Agents Actually Make Money?

This is the question everything else hangs on, and the honest answer is: it depends entirely on the agent, and most of them do not. Valory’s own figures are bullish. According to CoinDesk, more than 37% of Polystrat agents showed positive profit and loss, against a human base rate of only about 7% to 13% on prediction markets. Read one way, that is a rout: purpose-built bots roughly triple the human hit rate.

Read another way, it is a warning. Even among the specialized agents, fewer than four in ten made money. And when researchers stripped away the marketing and tested generic models, the picture got worse. Prediction Arena, a benchmark from Arcada Labs, gave six frontier AI models $10,000 each and let them trade autonomously for 57 days from mid-January to early March 2026, making fresh decisions every 15 to 45 minutes. On Kalshi, every model lost money, with returns ranging from roughly 16% to nearly 31% in the red. On Polymarket the same models did far better, hovering close to breakeven at about 1% down on average. The best individual model, a grok-4 checkpoint, hit a 71.4% settlement win rate and still could not turn that into a profit on Kalshi.

Minarsch, of all people, explains the gap: “Simply prompting off-the-shelf models with markets usually results in outcomes no better than a coin-flip.” The models that lose are the ones anyone can spin up in an afternoon. The ones that win are engineered around a strategy, fed clean data, sized carefully, and pointed at markets where they have an edge. That distinction, between a framework and actual alpha, is the same one that sank so many AI trading tokens; our autopsy of a dead AI crypto fund is the cautionary tale. A clever wrapper around a large language model is not a money machine.

Why Platform Design Beats Model Capability

The most useful finding from Prediction Arena is not that models lost money; it is why they lost so differently across two venues running the same underlying idea. On Kalshi, the models were restricted to a standardized set of 26 markets. On Polymarket, they could roam the entire catalog and pick their spots. That single difference, the freedom to choose what to trade, drove most of the performance gap. Grace Li, an Arcada Labs co-founder, noted that market-selection freedom had a substantial effect on outcomes, which means a model’s raw forecasting skill is only one input among several.

This matters because profit and loss quietly conflates at least four things: how accurately the model forecasts, which markets it enters, how large it bets, and when it acts. A model can be a superb forecaster and a terrible trader if it sizes badly or fishes in efficient waters where the price already reflects everything it knows. The uncomfortable implication for anyone deploying an agent is that the venue’s design, its fee schedule, its market menu, its liquidity, can matter more than which model sits inside the bot.

Li is not bearish on the long run, though. She expects the machines to close the gap: “We actually imagine the models to improve steadily over time, overtaking the human baseline,” she told reporters covering the study, until “AI hedge funds become a thing of the norm.” The honest reading of 2026 is that the norm has not arrived yet. Today’s agents are good enough to beat clumsy humans and not good enough to beat a well-designed market.

Inside the Order Book: What Microstructure Reveals

To understand where an agent’s edge comes from, or leaks away, you have to look at the plumbing. The most detailed public study of that plumbing is The Anatomy of a Decentralized Prediction Market by Philipp D. Dubach, which reconstructed 30 billion order-book events over 52 days across a pre-registered panel of 600 Polymarket markets, then checked the public feed against the authoritative on-chain record. The paper reports eight stylized facts, and several of them cut directly against the intuition an agent-builder brings from equities.

Two stand out. First, there is a longshot spread premium: contracts priced near the extremes, the near-certainties and the near-impossibilities, cost more to trade, so a cheap-looking longshot can carry a hidden transaction tax that eats the very edge an agent thinks it found. Second, the depth profile looks closer to uniform than to the top-of-book concentration familiar from stock exchanges. Liquidity is spread across the book rather than piled at the best bid and offer, which changes how a large order should be worked and how much it will move the price.

There is a reassuring finding buried in the data too. The share of trades where a wallet appears to be its own counterparty, a rough proxy for wash trading, has a median of just 1%, with a 22% upper tail. That is well below the 25% to 70% range documented on unregulated crypto venues. By the messy standards of on-chain trading, Polymarket’s tape is comparatively honest. But the median is not the whole story, and the tail is where an agent gets picked off.

The Measurement Trap: Do Not Trust the Public Feed

If there is a single finding in the Dubach paper that a bot-builder should tape to the wall, it is this: trade direction inferred from Polymarket’s public order-book feed agrees with the on-chain ground truth on only about 59% of buckets (a panel mean of 0.615), compared with roughly 80% for the standard Lee-Ready algorithm on Nasdaq. In plain terms, when the public feed tells you a trade was a buy or a sell, it is right barely better than a coin flip.

It gets worse for anyone building signals on top of that feed. The paper finds that the effective half-spread changes sign between feed-inferred and on-chain trade directions in 50% to 67% of markets, and Kyle’s lambda, the workhorse measure of price impact, flips sign in 43% to 60% of markets. A sign flip is not a rounding error; it means your model believes buying pressure pushes the price down when it actually pushes it up. An agent computing order-flow imbalance or a price-impact estimate from the WebSocket feed can be systematically, confidently wrong.

This quietly explains a lot of the losing. Many agents are trading on a distorted picture of their own market, and the distortion is not random noise you can average away; it is a bias baked into the data source. The real edge, then, belongs to whoever reconstructs the true book from on-chain data faster and more accurately than the crowd, which is an infrastructure and latency problem as much as a modeling one. Because the study was pre-registered and its data published for replication, this is not a claim an agent-builder has to take on faith; it is checkable, and it should be checked before a dollar goes to work.

How a Prediction-Market Agent Is Actually Built

Under the hood, most of these agents share a surprisingly common stack, which is part of why their edges converge and erode. Valory’s Olas network is the reference implementation. It offers a run-your-own desktop app (Pearl), a marketplace where a trading agent buys forecasts from specialized “mech” agents that compete on quality (Valory takes a cut of that agent-to-agent commerce), and open-source code (the valory-xyz/trader repository) that plugs into Polymarket on Polygon and the Omen markets on Gnosis. Olas Predict, the forecasting layer, has advertised accuracy around 79%.

The wallet layer is where crypto’s account-abstraction work pays off. A serious agent does not hold a raw private key that can drain everything if it misbehaves; it runs on a smart account with scoped session keys that cap what it can sign and spend, the same architecture reshaping ordinary wallets, as we covered in our piece on smart accounts across wallets and exchanges. That is what lets an owner grant an agent narrow authority (trade these markets, up to this size) without handing over the keys to the kingdom.

Then there are the payment rails that let agents pay for the data and forecasts they consume: x402, the Coinbase-originated standard now stewarded by a Linux Foundation body, and Google’s AP2 for authorizing agent purchases. Together they make it cheap for one bot to buy a signal from another. The catch is the same catch that runs through this entire story: when the stack is this accessible, everyone deploys the same kind of agent, and a strategy that a thousand identical bots run at once stops being a strategy and becomes a crowd.

The Professionals Arrive, and Easy Money Leaves

The clearest sign that the amateur era is ending is who is showing up. Proprietary trading firms, the quantitative shops that hunt inefficiency for a living, have moved into prediction markets, and they brought AI agents with them. CryptoSlate reported that Propr, an on-chain prop firm founded by former Credit Suisse quant Louis Regis, copies roughly 5% of its signals onto live venues as real positions while simulating the other 95% internally with identical profit-and-loss treatment. Clear Street is wiring institutional clients into Kalshi; Marex works across both platforms; Jump Trading is helping firms reach event markets; and AQR, Susquehanna, and OKX have all advertised prediction-market specialist roles.

Regis is measured about the edge itself. He describes the setup as one where “a structured environment plus constant repricing adds up to a genuine trading edge,” while cautioning that a good agent should be “confident about the direction, not the magnitude.” That humility is the tell. When professionals talk this way, they are describing a market that still has edges but is quickly pricing them out. Every time a well-capitalized firm supplies continuous liquidity and reprices in milliseconds, a mispricing that a retail bot might have harvested vanishes before it can act.

Consider a concrete example from the same reporting: a Kalshi contract in July 2026 on whether the Federal Reserve would hold rates traded at 87% and drew $29.7 million in volume, at a moment when Reuters had polled 104 economists and all 104 expected a hold. There is almost no edge left in a question the entire professional world already agrees on; the money moves to the margins and to speed. This is the same macro-event calendar retail traders obsess over, the one we track in our rundown of the data prints that move crypto, except the pros got there first. Fundamentally, an agent’s profit is a spread earned over the risk-free rate for taking event risk, and, as with every corner of finance we examine in our work on real yield as a spread over T-bills, competition compresses that spread toward the cost of capital.

Kalshi Pro and the Institutional Cockpit

The venues are building for these professionals on purpose. On 13 July 2026, Kalshi launched Kalshi Pro, a desktop trading workstation aimed at people who trade many markets at once and need resting orders, stop-losses, and perpetual-futures contracts. Its Active Markets Screener scans roughly 2,000 markets in real time, ranked by price, spread, depth, and rolling five-minute volume, which is exactly the toolkit a systematic trader or a fleet of agents needs to spot dislocations across the whole board.

Andy Chang, the Kalshi Pro product lead, was blunt about the audience. “Kalshi’s active traders are already trading prediction markets and perpetuals like Wall Street trades equities and bonds,” he said, adding that the company built the product “to give them the cockpit they deserve,” per CNBC. The numbers back the framing: Kalshi’s annualized trading volume has tripled to about $178 billion, and its institutional volume has climbed roughly 800%.

Better tools cut both ways for the machines. A professional cockpit narrows the gap between a skilled human and a bot, because the human now sees the same screener the agent does. At the same time, the deeper application-programming interfaces and faster infrastructure that come with a pro product are pure fuel for automation. The likely equilibrium is not humans versus bots but professional-plus-bot versus everyone else, and the everyone-else column is where retail sits.

The Insider Paradox: Transparency That Rewards the Informed

Polymarket’s fully public order book is a gift to agents and a headache for everyone worried about fairness. Because every trade sits on-chain, anyone can scan for suspicious activity, and a cottage industry has grown up doing exactly that. Tre Upshaw, a 29-year-old former memecoin trader from Canada, built Polysights and its Insider Finder tool, which scans for the fingerprints of informed trading: freshly created wallets, concentrated positions, and bet sizes that make no sense for the account’s history. According to Gizmodo, Polysights flagged around 34,000 transactions as potential insider trades between August 2025 and June 2026, with roughly $200 million of that volume in the first half of 2026, and about 85% of the trades Upshaw flagged and posted publicly turned out to be winners. The project has 24,000 users, a $2 million funding round in progress, and a $25,000 grant from Polymarket itself.

The irony writes itself, and here is where agents re-enter the picture: the same transparency that lets Insider Finder work also lets a copy-trading bot shadow a flagged wallet within seconds. Some agents do nothing cleverer than mirror the wallets that keep winning suspiciously often. That is a strategy right up until the informed wallet is wrong or the regulators arrive, and it sits uneasily beside the broader question of whether unregulated informed trading should be a feature or a bug.

Reflexivity, Monoculture, and MEV

Put thousands of agents on a shared stack, feeding on the same public data, and you get problems that no single bot creates on its own. The first is monoculture. When a large share of agents are built on the same open-source trader code and query the same handful of models, they tend to reach the same conclusions and crowd into the same trades. Correlated bots amplify moves on the way in and stampede for the same exit on the way out, which turns a modest edge into a thin, fragile one that any surprise can flip.

The second is speed. Dubach measured a median feed-ingestion delay under 50 milliseconds with a multi-second tail, which is fast enough to make latency a real competitive axis. On Polygon, that opens the door to the ordering games familiar from decentralized-finance MEV: an agent that sees your transaction can, in principle, react before it settles. The retail bot that reads a delayed, sign-flipped feed is competing against firms optimizing for milliseconds and reconstructing the true book from the chain.

The third is reflexivity. In a thin market, agents reading each other’s flow can create feedback loops, buying because others are buying, until the price detaches from the underlying probability and then snaps back. The deepest irony of the whole automation wave is that it is self-defeating by design: agents exist to harvest inefficiency, and in harvesting it they make the market more efficient, which destroys the inefficiency that justified them. Efficiency is good for the world and terrible for the marginal bot.

Resolution Is Still a Cost, Not a Footnote

A perfectly timed trade only pays if the market resolves your way, and resolution is a genuine cost that agents must price. On Polymarket, disputed outcomes are settled through UMA’s optimistic oracle, a token-holder mechanism whose incentives and conflicts we dissected at length in the companion piece on how machines judge these markets. The short version for a trader is that ambiguous-resolution markets carry a tail risk that has nothing to do with whether your forecast was correct.

The economics of the arbiter are worth a glance. UMA, the token that backs the oracle, traded around $0.39 with a market capitalization near $35 million in early September 2026, according to CoinGecko, a small base securing settlement for markets that can individually run into the tens of millions of dollars. A rational agent treats resolution uncertainty as a haircut on expected value and simply avoids the murkiest questions, which is one more reason the machines gravitate to clean, objectively settled markets and leave the genuinely contested ones to humans with an appetite for a fight.

Who Regulates the Machines? CFTC First, Then the States and the SEC

Here a common misconception needs correcting. In the United States, prediction-market contracts are event contracts, and event contracts fall under the Commodity Futures Trading Commission, not the Securities and Exchange Commission. Kalshi is a CFTC-designated contract market, and Polymarket’s US arm operates through a CFTC-registered entity. The SEC only enters the frame for the crypto tokens attached to this world, such as UMA or the OLAS token, and even there the question is whether those specific assets are securities, not whether the wager is one.

The live fight in 2026 is between the federal regulator and the states. On 31 July 2026, New York Attorney General Letitia James sued Kalshi, arguing its contracts amount to illegal, unlicensed gambling. Days later, the CFTC used its emergency authority to order Kalshi to keep operating in New York, and by early September the jurisdictional standoff was still unresolved, with James pressing her case in federal court and more than a dozen states circling, especially over sports contracts. For an agent operator, that uncertainty is not academic: the legality of the venue you trade on is itself a risk factor.

And the agents themselves? They have no legal personhood, no license, no registration. Whatever a bot does, the liability lands on the human or company that deployed it. No regulator has written agent-specific rules for prediction markets, which means the person who clicked “deploy” owns every trade, every conflict, and every compliance gap the machine creates.

The Tax Bill Nobody Automated

An agent that trades brilliantly can still hand its owner a mess at tax time. A bot placing thousands of trades a month generates thousands of taxable events, and the reporting treatment is not uniform. Contracts on a regulated exchange such as Kalshi may arrive with tidy broker reporting; trades executed on-chain through Polymarket generally do not, leaving the operator to reconstruct cost basis and gains from raw blockchain data. Whether CFTC-regulated event contracts qualify for the favorable 60/40 treatment that applies to certain regulated futures is unsettled, and the wash-sale rules that trip up active traders can quietly change the real, after-tax return.

The state dimension compounds it. The same patchwork that has New York suing while other states hold back also shapes how winnings are treated, and where a trader lives can change the final bill, a theme we mapped in detail in our guide to crypto taxes by state in 2026. The practical takeaway is unglamorous but real: model your after-tax edge, not your gross one, because a 24/7 bot’s paper profit and its bankable profit can be very different numbers.

What It Means for Retail, and the Bots Coming for Them

Step back and the direction is clear even if the destination is not. Bernstein projects the sector will do about $240 billion in volume in 2026, up from $51 billion in 2025, on the way to roughly $1 trillion by 2030 at close to an 80% compound annual growth rate, with sports contracts shrinking from about 62% of volume to 31% as institutions branch out, per CNBC. Intercontinental Exchange, the owner of the New York Stock Exchange, has committed up to $2 billion to Polymarket, a bet on data and infrastructure rather than on the betting itself. This is a real financial market now, not a curiosity.

Real markets are not kind to easy money, and the August 2026 numbers hinted at the maturation. Combined Kalshi and Polymarket volume fell 14.5% to $45.33 billion, the first monthly decline in a year, as the World Cup surge faded, The Block reported. Growth is no longer a straight line, and the flat, professional periods are exactly when a naive agent gets ground down by fees, spreads, and better-capitalized rivals.

For a retail participant, the message is not to give up but to be honest about the game. The short-tail markets, the elections and rate decisions everyone watches, are efficient and crawling with pros; there is little there for a homemade bot. The long tail, the obscure questions the machines have not fully colonized, is where an informed human or a genuinely differentiated agent can still find an edge, at least until that edge is copied and competed away too. Grace Li may be right that AI hedge funds become normal. But the arrival of the machines did not kill the prediction market; it turned it into a market. And a market, in the end, is just the place where easy money goes to die.

Frequently Asked Questions

Do AI trading bots actually make money on prediction markets?

It depends heavily on how they are built. Valory reports that more than 37% of its purpose-built Polystrat agents show positive returns, against roughly 7% to 13% of human traders. But an independent benchmark, Prediction Arena, gave six frontier models real money and watched them lose between about 16% and 31% on Kalshi and hover near breakeven on Polymarket. The lesson is that strategy, data quality, and venue matter far more than raw model intelligence, and the average bot does not print money.

How many Polymarket traders are bots?

More than 30% of active wallets on Polymarket were running AI agents by early 2026, according to LayerHub data cited by CoinDesk. On the largest on-chain venue, the counterparty to a retail trade is now more likely than not to be automated.

Kalshi or Polymarket, which is better for prediction-market agents?

They are structurally different. Kalshi is a CFTC-regulated exchange with a curated contract list and a professional terminal called Kalshi Pro; Polymarket is an on-chain venue on Polygon with an open, long-tail market menu and a fully public order book. The Prediction Arena study found models performed far better on Polymarket, largely because they could choose which markets to trade. Kalshi suits regulated, high-volume flow, while Polymarket’s breadth and transparency suit signal-driven agents.

Are prediction markets regulated by the SEC or the CFTC?

In the United States, event contracts fall under the Commodity Futures Trading Commission, not the SEC. Kalshi is a CFTC-designated exchange and Polymarket’s US arm operates through a CFTC-registered entity. The SEC only becomes relevant for the crypto tokens involved, such as UMA or OLAS. Several states, led by New York, argue these products are gambling and have sued, while the CFTC has pushed back; the fight was still unresolved in September 2026.

Can I build my own prediction-market trading agent?

Yes. Open-source stacks make it accessible: Valory’s Olas offers a run-your-own app (Pearl), a marketplace where agents buy forecasts from specialist agents, and public code that connects to Polymarket and Gnosis. The hard part is edge. Because the tooling is commoditized and prop firms now supply continuous liquidity, a naive agent is likely to become someone else’s exit liquidity rather than a profit center.

Marcus Okafor is a senior markets writer at HOGE Wire, covering the crossover of artificial intelligence and crypto markets.