Machine Politics: Bots Are Setting the 2026 Midterm Odds

On the first Tuesday in November, tens of millions of Americans will vote on who controls Congress. Weeks before a single ballot was counted, another electorate had already made up its mind: software. Autonomous trading agents, running day and night on Polymarket and Kalshi, have spent the autumn pricing every competitive Senate seat, every plausible House majority and every marquee governor’s race in real time. The probabilities they produce are quoted on cable news, inside campaign war rooms and across social feeds as if they were early returns coming in.

The 2026 midterms have widely been described as the first prediction-market election, and the phrase is more literal than it sounds. The odds that commentators treat as a read on the national mood are, more and more, set by machines trading against other machines. That changes what those odds mean, who can move them, and how much weight anyone should put on a number a bot printed at three in the morning. This is the story of that shift, told through the one event built to test it in public: a national election whose scoreboard is machine-run.

The first election the machines will call

Prediction markets are not new, and neither is election betting. What is new is the scale and the speed. In the week ending October 11, the two dominant venues together cleared a record $23.9 billion in notional volume, a fourth straight weekly high, with Kalshi taking roughly 71% of one Sunday’s $5.54 billion. Sports drives most of that raw turnover, but politics is the category that gets cited, and the midterms are the single largest political event either platform has ever hosted.

The growth curve is steep. Since the final weeks of summer, Kalshi’s weekly volume has climbed around 86% to roughly $19 billion, while Polymarket’s has risen about 85% to near $3.7 billion, both feeding that fourth straight record. A contested national vote is the rare event that pulls in sports bettors, political junkies, professional hedgers and, above all, the agents that never log off. Two years ago, during the 2024 presidential race, Polymarket’s Trump line became a story in its own right, repeated by anchors who had never heard of an optimistic oracle.

In 2026 the markets are deeper, the liquidity is larger and, crucially, the hands on the order book are mostly not hands at all. The odds are not a crowd of partisans and hobbyists typing in bets. They are the output of a population of trading agents, many built on the same handful of open frameworks, reading the same news wires and polling feeds, and repricing the same contracts within milliseconds of one another. The result looks like a consensus of the informed. It is closer to a machine average, and a machine average has very different failure modes than a human crowd.

What the odds say right now

As of mid-October, the machine consensus has swung hard in one direction. Through the first half of the year, Republicans were favored to keep the Senate; by the autumn that had reversed. Traders on both Kalshi and Polymarket now favor Democrats to take the chamber, a move from Republican odds near two-thirds in January to roughly one-third by early October. The House, where the party holding the White House almost always loses ground at a midterm, is priced as a near lock to flip.

The snapshot below reflects Polymarket’s main congressional markets in mid-October, drawn from on-chain data compiled by Bitquery. These are live markets: the numbers move intraday and will keep moving until they resolve on November 3. Read them as a picture of where the machines sat on one October afternoon, not a forecast carved in stone.

Notice the volumes. For all the attention they draw, the dollars riding on who controls the Senate are modest beside a single NFL weekend. That mismatch, a thin market carrying an outsized share of the national conversation, is the first thing to hold in mind about how much weight these odds can really bear. A few million dollars decides a number that millions of people read as destiny.

The electorate behind the odds is mostly software

Who is actually trading? On Polymarket, the answer is increasingly not a who at all. By analytics reported through CoinDesk earlier this year, more than 30% of active wallets were running AI agents, and 14 of the top 20 wallets by activity were bots rather than people. The agents are also better at it: roughly 37% of them finished in profit, against a human hit rate down in the low teens. When a market is dominated by participants who do not sleep, the baseline a human trader competes against is no longer another human.

David Minarsch, chief executive of Valory, the company behind the Olas agent stack and a trading agent called Polystrat, describes the product plainly. It is, he told CoinDesk, “an autonomous AI agent that trades on Polymarket 24/7 on behalf of its human user.” His caveat matters just as much: simply prompting an off-the-shelf model with a market, he has said, usually yields results “no better than a coin-flip.” The edge, where it exists, comes from engineering around the model, not from the model alone. Even so, Minarsch argues, agents tend to beat humans, partly because people make rushed decisions and machines never get bored, distracted or emotionally attached to a candidate.

Election markets are a natural home for this kind of automation. They run for months, they resolve against a single public fact, and they throw off a constant stream of polls, endorsements and headlines that a model can digest faster than any person. The table below sketches why the political book suits agents, and where that fit quietly breaks down.

How you build an agent that trades a Senate race

The barrier to entry has collapsed. Polymarket maintains an open-source agents framework on GitHub, essentially a starter kit for wiring a language model into live markets, and third-party trackers now list dozens of ready-made politics-and-elections bots that trade on polls, news and order-flow momentum. None of this requires a trading desk or a quant team. A motivated developer with an API key and a funded wallet can field an election bot in an afternoon.

Under the hood, an election agent is three parts. There is a brain, usually a frontier language model, that reads inputs and forms a probability. There is a wallet, so the agent can sign its own trades, typically funded with stablecoins and topped up over machine-payment rails. And there are hands: the code that places orders, sizes positions and takes profit. Some builders go a step further, hiring specialist forecasting agents through a marketplace so the trader does not have to grow its own model from scratch, an early and still rare example of machines paying machines for finished work.

That division of labor is one of the few places the much-hyped machine economy actually runs in production. A trading agent can outsource the hard part, forming the probability, to a specialist forecasting service and simply act on the answer, paying for it in stablecoins over the same rails. It is a glimpse of agents contracting other agents for a living, and prediction markets are where it first became routine rather than a demo. The flip side is homogeneity: when many traders buy their edge from the same few services, their positions start to rhyme, and a market that looks like a thousand independent opinions can be a handful of models wearing a thousand masks.

The flaw in this design is the one that haunts every autonomous agent: the model cannot reliably tell instruction from data. An agent scraping a social feed for sentiment can be fed a planted headline, a fabricated poll or a poisoned link and act on it before any human notices. In a market priced by dozens of agents reading the same sources, one convincing piece of disinformation can ripple through the order book in minutes. The order-flow games bots play against each other are close cousins of the sandwich and front-running tactics that plague decentralized exchanges, which HOGE Wire has examined in the context of trading bots and the law.

The 2024 cautionary tale: the whale who moved the odds

To see why a thin, fast market is fragile, look at the last big one. In the closing weeks of the 2024 presidential race, a cluster of four Polymarket accounts, later traced to a single French trader the press nicknamed Theo, poured tens of millions of dollars into bets on Donald Trump. One account alone, Fredi9999, held more than $20 million in pro-Trump positions. The buying was heavy enough that one market, on whether Trump would win the popular vote, leapt from the mid-20s into the high-30s in percentage terms within hours.

Polymarket investigated and said it found no evidence of manipulation, noting the trader had deliberately spread his orders to avoid jolting the price and had described the wager as a personal conviction call. He was also right: blockchain-analytics firm Chainalysis later estimated he netted close to $79 million. But the episode laid bare the structural problem. A single well-capitalized actor could visibly tilt the headline odds of a US presidential election, and the public had no way to know, in the moment, whether the move reflected real information or just a large bankroll with a view.

Now replay that with agents. A bot does not need conviction; it needs a thesis and a bankroll. On a down-ballot Senate race carrying only a few million dollars of liquidity, the Theo problem shrinks to a fraction of the capital it took in 2024. The machines that dominate the order book are precisely the actors best equipped to shove a thinly traded political contract, and to do it faster than any human watchdog can blink. The 2024 whale was a person with a hunch. The 2026 version could be a script with a prompt, repeated across a dozen wallets.

Do the markets actually know who will win?

Boosters like to say the odds beat the polls. The record is messier. In 2024 the markets did lean toward Trump while polling averages showed a coin flip, and that looked prescient when he won. But a careful postmortem complicates the victory lap. A 2025 academic review of roughly $2.4 billion in 2024 election trading found the markets tracked the national presidential race about as well as ordinary statistical models, and did noticeably worse on down-ballot congressional contests.

That is the uncomfortable detail for this cycle, because a midterm is nothing but down-ballot congressional contests, the exact races where the markets stumbled last time. Accuracy also swings wildly by venue; across the studies, the more regulated, lower-stakes books have often scored better on calibration than the deepest-liquidity crypto venue, a reminder that more money does not automatically buy more wisdom. Polymarket’s surge of 2024 liquidity may even have worked against it, drawing in more speculative noise alongside the signal.

There is a subtler problem on top of that. Frontier AI forecasters, when tested against real market questions, now roughly match the market on calibration but stay systematically too cautious, rarely committing to the extreme, high-conviction probabilities that actually win money. If the agents collectively setting an election line are all hedging toward the middle, a market that looks confident may be masking a crowd of models that privately are not. Calibration is not conviction, and an election is decided in the tails, not at the average.

The honest summary is that prediction markets are a useful signal and a poor oracle. They aggregate real information, react quickly, and are harder to fool than any single poll. They are also thin where it counts, gameable at the margins, and prone to projecting more certainty than they have earned. For a midterm fought across dozens of close, low-liquidity races, that combination argues for reading the odds as one input among many, not as the result leaking out early. The number is informative. It is not the same as knowing.

When the odds become the news

The most consequential thing about a machine-set election line is not whether it is right. It is that people treat it as if it were. A price on a screen is a single, vivid number, and it travels. When a respected Iowa poll late in 2024 showed the race tightening, candidate odds on the betting markets moved in direct response, and that move itself became a headline. The market reads the news, the news then reports the market, and the market reacts to its own coverage, a loop with no obvious brake.

Economists call this reflexivity, and prediction markets are an unusually potent case because the price does something a poll never does: it collapses a cloud of scattered private beliefs into one salient focal point that everyone can see and cite. On an election, that is not an academic worry. Election officials have warned that a bot-driven odds line flashing ninety-something percent for one side could be dropped into misleading social posts to imply a race is already decided, discouraging people from turning out before a single polling place has closed.

This is where the automation and the influence compound. The odds are produced by machines optimizing for profit, not for civic accuracy. They are then laundered through coverage into something that looks like a verdict. Nobody at a trading agent set out to shape turnout, but a number that moves money and moves headlines can, at the margin, move voters too, and no bot was ever asked to care about that side effect. The market was built to aggregate belief; it has become a machine for manufacturing it.

Policing a market that never closes

Both platforms know the integrity question is now existential, and both have moved. Polymarket has built out a market-integrity team under Shauna Bautista, a former FBI investigator. Ahead of November she told PYMNTS, “I can tell you that we have the systems in place to be able to identify anomalous activity when the midterms do come,” leaning on machine learning, blockchain analytics, trade surveillance and open-source research. The defense against predatory bots is, inevitably, more bots.

Detection is not prevention, though, and the two venues start from very different places. Kalshi is a federally regulated exchange that knows its customers; every account is identity-verified, which makes an insider or a manipulator traceable. Polymarket’s main platform is offshore and pseudonymous, which is exactly how the Theo accounts could move a presidential market from behind four nicknames. That contrast between a know-everyone venue and a know-no-one venue mirrors the broader fault line in crypto compliance that HOGE Wire mapped in its look at the two camps of DeFi.

Insider trading on political contracts is no longer hypothetical. Over the past year, federal prosecutors and the Commodity Futures Trading Commission brought their first cases against traders who acted on non-public information in event contracts, and the regulator’s enforcement chief has publicly rejected the idea that insider trading is somehow tolerated on prediction markets. The table below sets the 2024 integrity picture against where things stand for the 2026 vote.

Who calls the race, and what if it is close

Every bet has to be settled, and settlement is the most brittle joint in a machine-run election market. The two platforms do it differently. Kalshi, as a regulated exchange, resolves against named source agencies filed with its regulator, the same kind of authorities that call races on election night. Polymarket resolves through UMA, a decentralized optimistic oracle: someone proposes the outcome, a short challenge window opens, and if anyone disputes it, UMA token holders vote to decide the truth.

That works cleanly when the result is obvious. It does not when a race sits inside a recount margin, when a call is delayed for days, or when who won is being argued in court. An oracle built to resolve in hours is poorly matched to an election that can take a week to certify, and a token-holder vote on who won a Senate seat is a governance surface nobody should want to stress-test in public. The industry’s proposed fix is to automate resolution too, handing the call to AI judges locked to a fixed model and prompt at the moment a market is created, an approach that rhymes with the verifiable-AI ideas HOGE Wire covered in its piece on proving a model’s output on-chain. A perfect proof of what a model said is still only as trustworthy as the model that said it.

Picture a Senate seat called on a Friday, then contested in court by the following Tuesday. Kalshi can lean on the agency its contract named and settle. Polymarket’s oracle has to decide whether the race was decided, a question that is itself the subject of litigation. Whichever way the token vote lands, half the market will insist the oracle got it wrong, and because the vote is public and on-chain, the dispute becomes part of the permanent record of that election. A settlement layer built for sports scores and crypto prices was never meant to adjudicate contested democracy.

For the midterms the danger is concrete. Most races will resolve without drama. But the handful that are genuinely close, the ones carrying the most money and the most attention, are exactly the ones where a resolution dispute could turn a prediction market into a second, parallel fight over the result, playing out days after the polls close and feeding straight back into the public argument about who really won.

Are election bets even legal?

The machines are trading into a legal vacuum that Washington has spent two years failing to fill. Election contracts were effectively off-limits until recently. The CFTC moved to block Kalshi from listing congressional-control markets, Kalshi sued, and in 2024 a federal judge ruled the agency had exceeded its statutory authority. An appeals court let the contracts stand, and Kalshi became the first federally regulated venue in more than a century to offer legal election trading in the United States.

The fight did not end there; it spread. On October 9 the CFTC proposed formally classifying sports, political, cultural and weather event contracts as swaps, while issuing a separate interim rule carving state-licensed casino and sportsbook wagers out of that definition. The swaps piece is only a proposal, not a final rule. Meanwhile three federal appeals circuits have split on whether states can regulate these markets at all, New Jersey has asked the Supreme Court to weigh in, Kalshi’s response is due in early November, and the justices are not expected to decide whether to hear the case before December. HOGE Wire tracks that timeline in its running regulatory countdown.

One framing point cuts through the noise. Prediction-market contracts are event contracts, which fall under the CFTC and the Commodity Exchange Act, not under the Securities and Exchange Commission. The SEC enters only at the edges: the crypto tokens this infrastructure runs on, and any actual stocks an agent might buy on the back of a market signal. For the bets themselves, the relevant US regulator is the CFTC, and the open question is whether even it gets the final word over the states.

Follow the money, and the politics

There is one more reason the legal question is so charged: the people who stand to gain are close to power. Donald Trump Jr. has become a central figure in the business. His venture firm, 1789 Capital, led Polymarket’s latest funding round, and he advises both Polymarket and Kalshi for pay. Kalshi, for its part, has spread its political bets, steering money into state-level political action committees after unfriendly hearings and hiring former Biden administration officials as a hedge against whichever way the wind blows.

Kate Knibbs, a senior writer at Wired who covers the sector, put the conflict directly to The Intercept: “Donald Trump Jr. is very involved in this industry, and stands to gain a lot from prediction markets being able to operate as they desire.” When the family of a sitting president holds a financial stake in whether a market that prices that president’s own party stays legal, the regulatory debate stops being a purely technical argument about the Commodity Exchange Act.

None of this is illegal, and that is the point. The conflicts are structural, disclosed and perfectly lawful, which is exactly why they are hard to police. A venue can run a scrupulous market-integrity program and still be owned by people with an enormous interest in which way the regulatory and electoral winds blow. For traders, the takeaway is simpler than the politics: the odds you read are produced inside an institution that is not a neutral referee, and has never claimed to be one.

This is the part the clean probability on a screen never shows. Behind it sits an ownership structure, a lobbying budget and a set of incentives that have nothing to do with forecasting an election and everything to do with being allowed to keep running the book. The agents trade; the humans who own the venues play a slower, larger game about the rules of the game itself.

The tokens nobody is betting on

For all the money sluicing through these markets, the crypto tokens attached to the plumbing have been left behind. UMA, the token that secures Polymarket’s resolution layer, trades around $0.39 for a market value near $37 million, a rounding error next to the billions in weekly volume it helps settle. The Olas token that anchors one of the most-used agent frameworks carries a market value in the low tens of millions. The agents print real volume; the tokens capture almost none of it.

It is a lesson the AI-crypto sector keeps relearning: a useful framework is not the same thing as a valuable token. The best election-trading stack and the oracle that settles its bets can both be indispensable while their tokens drift, because the value accrues to the platforms, the whales and the agent operators, not to a governance coin most traders never hold. The gap between usefulness and price is a running theme in this corner of crypto, and it is especially stark here, where the infrastructure is busier than it has ever been and the tokens are quieter than ever.

That split also marks where the SEC could ever get a foot in the door. The bets are the CFTC’s turf, but the tokens are financial assets in their own right, and an agent that buys a stock or a crypto asset on the strength of an election-market signal drags the activity back toward securities law. The machine that reads a Senate line and then loads up on a defense contractor has quietly crossed from one regulator’s map onto another’s, often in the same second.

What to watch on and after November 3

The vote itself is the easy part. The real test for machine-run election markets comes in the hours and days around it. Watch three things. First, resolution: do the close races settle cleanly, or does a delayed or disputed call drag the oracle problem into the open? Second, reflexivity: does the press keep quoting bot-set odds as results trickle in, and does that coverage visibly push the markets around? Third, integrity: do the new surveillance programs actually catch anything, and does any late, suspicious lurch in a thin down-ballot market look like the Theo playbook, this time run by software?

Beyond election night, the calendar is crowded. The Supreme Court’s decision on whether to take up the states-versus-CFTC fight is expected to come into focus around December, the same month the Federal Reserve’s next rate move will be priced on these very platforms, a market HOGE Wire follows in its outlook on what rides on December. Prediction markets have quietly become the place where the market’s own expectations about politics and policy get set, which makes their integrity a public question, not merely a trader’s.

There is a larger experiment running underneath the midterms. If machine-run markets handle a messy, high-stakes national election without a manipulation scandal or a resolution fiasco, the case that prediction markets are serious financial infrastructure, rather than a casino in a civic costume, gets much stronger, and the institutional money already circling grows bolder. If they stumble in public, the backlash will not land on the bots. It will land on the platforms, their politically connected backers and the regulator that let the whole thing scale. November is a test the machines did not ask to sit.

The deepest change is the quietest. For most of modern history, the single number that summarized an election before the votes were counted came from pollsters with methods you could argue about. In 2026 that number increasingly comes from a swarm of trading agents optimizing for profit, settled by an oracle, and owned by people with a direct stake in the fight over their own legality. The machines are not so much forecasting the election as pricing it, and the rest of us are finding out, in real time, how much that distinction matters.

Frequently Asked Questions

Are AI agents really setting the odds on the 2026 midterms?

In large part, yes. On Polymarket, more than 30% of active wallets were running AI agents and 14 of the top 20 wallets by activity were bots, according to analytics reported by CoinDesk. These agents trade around the clock on polls, news and order flow, so the headline odds that media outlets quote are shaped heavily by software rather than by individual human bettors.

Are prediction markets more accurate than polls for elections?

Not reliably. Markets leaned toward the eventual winner in the 2024 presidential race, but a 2025 academic review found they tracked the national result about as well as ordinary statistical models and did worse on down-ballot congressional races, which is exactly what a midterm is. Accuracy varies a lot by platform and by how much liquidity a given market carries.

Is it legal to bet on US elections on Kalshi or Polymarket?

Kalshi, a CFTC-regulated exchange, won a 2024 court ruling that let it offer legal US election contracts, the first regulated venue to do so in over a century. Polymarket’s main platform is offshore and officially off-limits to US users. The law is still unsettled: the CFTC has proposed new rules, federal appeals courts have split, and the Supreme Court may yet weigh in.

Can someone manipulate an election prediction market?

It has already happened. In 2024 a single French trader using four accounts moved a Polymarket presidential market sharply with tens of millions of dollars in bets. Thinly traded down-ballot races are even easier to push, and AI agents are well equipped to do it quickly. Platforms now run machine-learning surveillance to flag anomalous activity, but detection is not the same as prevention.

Which regulator oversees prediction markets, the SEC or the CFTC?

Prediction-market contracts are event contracts under the Commodity Exchange Act, which puts them under the CFTC, not the SEC. The SEC becomes relevant only at the edges, such as the crypto tokens the infrastructure uses or any stocks an agent might trade based on a market signal.

By Marcus Okafor, senior markets writer at HOGE Wire, covering the collision of AI agents and crypto markets.