2026-11-11 / essays / 6 min read

AI Agents in Prediction Markets: Accuracy, Manipulation, Disputes

AI agents already trade prediction markets through Polymarket's SDKs and agent skills, Olas's Polystrat and Public's Kalshi tie-up, and machine-run wallets placed 63% of Polymarket trades in a year to August 2026. Our view is that agents improve deep, numeric markets and raise manipulation and dispute risk in thin markets that settle on wording.

Our position is that AI agents will make prediction-market prices more accurate on deep markets that settle on a published number, and less trustworthy on thin markets that settle on wording, because agents cut the cost of trading, of quoting and of arguing a resolution by the same amount, and only the first two are checked by other traders.

What is shipping now?

The tooling for agents to trade is public and mostly official. Polymarket's GitHub, read on November 11, 2026, shows the shift. The original Polymarket/agents framework ("Trade autonomously on Polymarket using AI Agents", 3,817 stars) is archived, with its last push on November 5, 2024. The old py-clob-client is archived too; its README says it "is no longer functional." Their replacements are unified Python and TypeScript SDKs, both pushed on November 11. An agent skill published on February 19, 2026 "gives agents the knowledge to authenticate, place orders, read markets, stream real-time data, manage positions, bridge assets across chains." The Polymarket CLI (2,907 stars) offers a JSON API "for scripts and agents," with a warning that it is "early, experimental software."

Kalshi has drawn a narrower line. Its hosted MCP connector gives an AI assistant 22 tools for markets, perps, portfolio and price alerts, and "does not place or cancel orders, transfer funds, or save trade drafts." Execution by agents on Kalshi comes through a broker: Fortune reported on September 24 that Public's agents turn a user's instructions into fixed rules, and "once activated, the agent cannot make trades outside the customer's chosen conditions." Kalshi provides the contracts and handles the trades.

Third parties fill the rest. Olas launched Polystrat on February 9, 2026, an agent that runs "via Pearl using self-custodial Safe accounts" and takes strategy "instructions in plain English."

How much trading is already automated?

More than the agent products suggest. Bitquery's onchain study of 1.65 billion Polymarket trades from September 1, 2025 to August 19, 2026 classed a wallet as machine-run if it placed orders in all 24 hours of some day and its median gap between orders was under 60 seconds. About 0.5% of wallets qualified. They placed 63.1% of trades, 28.3% of money traded and 36.6% of fees, and in August 2026, 49% of money had a machine on one side. Bitquery calls its tests floors. A timing rule catches market makers and scripts as well as language-model agents, so as a count of AI agents the figure is an upper bound.

The LLM-agent slice is small. Olas's dashboard, as of 21:20 UTC on November 11, 2026, reported a seven-day average of 17 daily active Polystrat agents on Polymarket and 141 registered in total. On Omen, it reported a seven-day average of 200 daily active trader agents and 16.8 million trader-agent transactions.

Do agents make prices more accurate?

The best published evidence says AI forecasters have caught up with skilled humans overall and have not yet beaten a market price. On ForecastBench's leaderboard as updated on November 11, 2026, the top entry scored 69.5 overall against 68.2 for the superforecaster median, a gap the board marks as not significant (p = 0.89). The benchmark also scores questions drawn from Kalshi, Polymarket, Metaculus, Manifold and INFER. On those, its "Naive Forecaster," which submits the market's value from the day before the forecast date, scored 70.6. The top model scored 71.5, with a 95% interval of 70.4 to 72.7, and the superforecaster median scored 73.8. The Prophet Arena paper found LLMs well calibrated but with "slower information aggregation compared to markets when resolution nears."

Trading results point the same way. Olas reported Polystrat's trading ROI at -7% since January 16, 2026, with 28.6% of 42 qualifying agents net positive over 90 days. Bitquery found 34.4% of machine-run wallets finished ahead on markets opened and settled from August 11 to 19, 2026, against 37.8% of other wallets.

Our read is that agents add accuracy mainly through speed and coverage: they keep many correlated contracts consistent and close gaps between venues faster than people can. That helps most where a price feed or official statistic settles the contract and many agents compete, so one agent's error is another's profit.

Where do agents raise manipulation risk?

In thin markets, where nobody is paid to trade against them. Our liquidity pull found 86.7% of contracts in Polymarket's 300 deepest events with no trades in 24 hours, and our wallet-concentration essay cited ACDC's finding that the top 1% of wallets produced 68% of Polymarket's 2026 congressional-market volume. An agent costs little to run across thousands of such contracts, and a modest budget can move a price nobody is defending. The cost that used to limit this was attention, and agents remove it.

Correlation is the second risk. CFTC Chairman Michael Selig said on October 28, 2026 that "systems relying on similar models may generate correlated trading behavior," and asked whether markets need "KYA or know-your-agent." His remarks did not mention event contracts. Agents built on the same few models and the same news feeds will tend to read a headline the same way at the same moment, which can look like consensus and is one reading copied many times.

The third risk is in contracts settled by a person's conduct. The CFTC's September 22, 2026 advisory said mention markets "present a heightened risk of manipulation." An agent that monitors a speaker in real time and trades the instant a word is said is fast; one that can also influence what gets said, or file the dispute about whether it counted, is a problem.

What do agents do to resolution disputes?

They lower the cost of disputing. An April 2026 study of UMA-resolved Polymarket markets counted $972,370,804.71 of volume on disputed events and found web-enabled LLMs reached "89.58% agreement with UMA's final resolutions" once a dispute was raised, while "unable to reliably predict which events will become disputed in advance." UMA's own Optimistic Truth Bot, in a June 3, 2025 post, ran at about 78% accuracy across market types, about 95% excluding 50/50 and too-early outcomes, and 72% on mention markets, and was "not yet making these proposals onchain."

Those numbers carry our concern. Agreement near 90% is good for a reviewer and dangerous for an automated disputer, because each of the remaining cases costs only a bond to contest, and an agent with a position has a reason to contest them. Our oracle essay cited a Wall Street Journal count, reported by Finance Magnates, of more than 1,150 disputed Polymarket markets in 2026 by mid-May, already past the 2025 total. We expect agents filing well-written disputes at machine speed to push that count up, concentrated on the same wording-dependent markets where the models are weakest.

Should agents create markets?

On Omen they already do: Olas counts 104,074 transactions by market-creator and closer agents. On Polymarket's new contracts, opening a resolution request is an operator-gated call, per the V2 oracle documentation, so only Polymarket's operator can open one. We think automated creation is fine for questions that settle on a data feed with fixed parameters, and should stay human-reviewed for anything whose rules a person has to interpret. Our Protocol V2 essay describes the authorized-reporter module that would let an operator plug an automated reporter into settlement.

What is the strongest case against our view?

The strongest objection is that agents make wording-dependent markets safer. A model that reads every rule and every clarification will spot ambiguity before money piles in, and an LLM that agrees with final outcomes 90% of the time could screen proposals before any bond is posted. Cheaper disputes mean more scrutiny, and scrutiny is what token-holder votes lack. Manipulation in thin markets also predates agents.

We agree that agents are useful reviewers, and the venue that runs one on its own rules before listing will write better contracts. Our claim concerns incentives. A reviewer paid by the venue and an agent paid by its position do the same reading and file different disputes. The thin-market point holds: agents did not invent manipulation, they made it cheap to run across ten thousand contracts at once.

What would change our mind

• Calibration data showing thin Polymarket contracts became more accurate after September 2025 as machine-run share rose.
• Dispute counts per market staying flat while automated trading grows, or LLM-screened proposals cutting disputes on mention and wording markets.
• An AI system beating the market price on ForecastBench's market questions by a margin outside the confidence interval.
• Evidence that venues can identify and cap correlated agent activity without banning automation, along the lines of the "know-your-agent" idea Selig raised.