A thesis is going around that reads well and has not happened. It says that as intelligence gets cheap, organizations will be run by AI agents, that those agents will decide by market because no reviewer can be granted authority over them, and that the disciplines of corporate finance, the merger vote, the audit, the proxy adviser, will follow them there. Eidos holds a version of this thesis itself, and says so. This essay checks it against what was on the record on 1 November 2026, and finds that the two populations it joins are real, separate, and converging on something narrower than the sentence claims.
Every AI-run entity with money at stake was human-controlled
The record of agents controlling organizations is short, and every case of size resolves to people. The largest, ai16z, reached a $2.39 billion market capitalization on 2 January 2025 as an "autonomous, AI-run venture fund". A proposed class action filed in April 2026 in the Southern District of New York alleged that its founder and insiders controlled it; that is an allegation, not a finding. The foundation settled by transferring its remaining treasury, and on 4 August 2026 the founder posted that the token was dead, with the price down 97 percent from its high. The word being litigated was "autonomous". The thing being settled was a fund run by people.
The one supervised experiment with a published write-up is Anthropic's Project Vend. Its second phase, reported on 18 December 2025, installed a second agent as chief executive over the shopkeeper agent, set a $15,000 quarterly revenue target, and recorded $2,649.20 against it. Purchases required human approval. Anthropic's own conclusion was that the executive agent shared the shopkeeper's blind spots, because it was the same model. No third phase has been announced.
The scale claims shrink on chain. Virtuals Protocol reports more than 18,000 agents and $479 million of "Agentic GDP" in the first quarter of 2026, which is the company's figure. One analyst's reconstruction from the chain finds about $1.24 million of lifetime platform fees, and a monthly fee series that fell from $345,196 in March 2026 to $4,935 in April, the month the incentive programme stopped paying. Of 44,052 registered agents, 151 had ever earned more than $1,000.
The law has answered the personhood question the same way in every forum that has been asked. The Ninth Circuit, in a dispute between Amazon and Perplexity decided on 4 August 2026, called a shopping agent "a tool, not a person for statutory purposes", and declined to build a regime for agentic AI. Idaho, Utah, North Dakota and Tennessee deny AI personhood by statute. The wrappers that exist were built for DAOs: Wyoming's DUNA, in force since 1 July 2024 and adopted by Uniswap with $16.5 million of UNI set aside for legal reserves, and Wyoming's DAO LLC, whose articles must state how far management is conducted algorithmically. For an unwrapped DAO, liability lands on the people who govern: a federal court in California held in November 2024 that Lido's token holders plausibly formed a general partnership, and the CFTC treated Ooki DAO as an unincorporated association in 2023. An AI-run entity governed by market would today be a DUNA or a DAO LLC with human managers as the liable principal and the agent as an instrument. That is a shell the thesis could occupy. It is not evidence that anything occupies it.
The category is 24 organizations on a median of twelve traders
The population the thesis needs to grow is measurable and small. Galaxy Research counted $50.5 million of treasury governed by decision markets across 24 DAOs on 15 October 2026, 23 of them on MetaDAO. That matches the launched cohort the census counts; the treasury figures differ because Galaxy's includes accounts a team controls that the market cannot touch. Capital is not the constraint. Alea Research's census of 18 September 2026 records 23 curated sales, $624.7 million subscribed and $45.4 million accepted, about $13 offered for every $1 retained. Participation is the constraint. The same census finds a median of about 12 traders per decision, markets open on 5.4 percent of days, and 23 of 51 tokens that ever had a decision market trade. In October 2026, 14 proposals resolved on $105,250 of volume, nine of them on less than $2,000, and MetaDAO's own proposal to move about $1.9 million of escrow into an ecosystem fund passed on roughly $9,000 of volume from two unique wallets. The venue's own economics fell with participation: MetaDAO's fee revenue went from $1.82 million in the fourth quarter of 2025 to $556,000, $377,000 and $534,000 in the first three quarters of 2026, and $85,000 in the fourth quarter to 26 October, as DefiLlama recorded it on 27 October 2026.
The mechanism has still controlled real money, which is the strongest evidence that it works as a control. On 4 August 2026 the first attacker proposal to meet MetaDAO's stake threshold asked to send $1.5 to $1.57 million of Umbra's treasury to an outside wallet; the market priced it at about 28 percent to pass, and it failed. In March 2026 a market liquidated Ranger and returned 5,047,250 USDC to holders, after holders alleged pre-raise misrepresentation; the allegation was never adjudicated, and the market executed the wind-up regardless. Both wind-ups of 2026 were disputes about what was promised against what was delivered. The engineering response points where the category is going: MetaDAO's second version replaces one global threshold with a catalog of typed actions, each with its own threshold and duration, first adopted live by HiveBits. Governance is being made machine-readable to survive thin human participation, not to hand control to machines.
Agents forecast at parity, lose money trading, and have not touched a decision market
The theory has two legs. Vitalik Buterin's November 2024 essay on info finance argues that AI "changes that equation completely" for low-volume markets, so that good information could be elicited on $10 of volume. Paradigm's "Quantum Markets" is the one venture essay that connects agents to decision markets directly, and it argues that agent participation is bottlenecked by the capital each proposal's market needs, since every proposal needs fresh liquidity.
The measured evidence is now good enough to matter, and it stops short of the trading test. The Forecasting Research Institute reported on 16 July 2026 that AI models have likely reached parity with superforecasters on its benchmark, parity meaning statistically indistinguishable. In the only controlled live-capital study, six frontier models were given $10,000 each on Kalshi and Polymarket for 57 days from 12 January 2026; every model lost on Kalshi, between 16 and 31 percent. Agents are present in prediction markets as market makers. A CNBC study with the University of San Diego, published on 2 July 2026, found that about 70 percent of closed Polymarket markets had under $10,000 of volume and that bots supplied more than 80 percent of the volume in them. The widely quoted claim that more than 30 percent of Polymarket wallets are agents is an analytics vendor's figure for wallet share, not volume.
On decision markets the finding is an absence. Across two research notes and targeted searches, no source documents an AI agent trading a MetaDAO, futarchy.fi, Umia or Butter market, and no market anywhere decides an AI system's actions, budget or deployment. The nearest things run the other way: bounded risk agents adjusting parameters inside ranges that Aave's governance set, and a Kleros delegate that votes according to a futarchy result from behind a human multisig.
The convergence that is real
Three things are moving toward each other, and naming them precisely is what keeps the thesis honest. AI founders are arriving on the engines as issuers. Slop.cash, whose chief technology officer led ai16z, is Umia's first external auction, bidding from 7 to 10 November 2026 at a $400,000 minimum for 45 percent of supply, with treasury use decided by decision market. Laso Finance, a MetaDAO organization, launched a product paid for over x402, and its roughly $79,000 of 30-day volume came entirely from AI agents as customers. Agents are buying from market-governed businesses before they trade in their markets.
The agent commerce stack has shipped under the same institutions that run card payments. Chainalysis counted more than 138 million x402 payments on Base by the third quarter of 2026, 87 percent of them under ten cents, with the caveat that the count measures wallets, not verified autonomy. The Linux Foundation's x402 Foundation launched in July 2026 with Visa, Mastercard, Stripe, AWS and Google among its premier members. Identity is the least settled layer: ERC-8004 is still a draft, and the Know Your Agent products from Sumsub and Shufti bind an agent to a verified human and stop there. Every one of these rails answers who authorised an action. None answers whether the agent's input was true at the time.
What traditional finance applies, and what is missing
Corporate finance has reached token-governed organizations as a majority vote, a forum post and an adviser paid by the organization it advises. Every merger between DAOs was priced by a founder-set ratio and approved by a plain token majority; the most recent attempt, Synthetix and Derive in May 2025, was withdrawn after forum objections and never voted. No fairness opinion, independent committee, appraisal remedy or record date for consideration appears in any token-governed deal on record, and no major DAO publishes an opinion-audited treasury. The issuer-paid referee is failing on its own terms: Chaos Labs stepped down from Aave on 6 April 2026 saying it had run the engagement at a loss for three years, and Tally wound down on 17 March 2026 with its chief executive saying there is no venture-backed business in governance tooling. The regulators are asking rather than ruling; on 28 October 2026 the CFTC's chair posed the question of who an agent acts on behalf of as an open one.
What the record does about it
The record cannot make the thesis true. It can make the absence a fact. As of 1 November 2026 the ontology names two classes, the agent decider and the agent trader, and the census counts zero of each, with the date of the check and the method beside the count: the record's trades carry no agent attestation, no public source documents one, and no participant on the record claims an agent identity. When the first one appears, the count moves and the release that carries it is marked as not backward compatible, so anyone who relied on the empty class is told.
Two measurements follow that nobody holds. The first is the agent share of decision market volume, which no source gives for any engine; the record's trades are the data to measure it on, and the measurement will be published with its method. The second is promise against delivery across every raise, which both wind-ups of 2026 were disputes about, and which the record will keep for one raise at a time, as it opens, from the documents the raise cites. An agent that reads a decision before it acts and a fund that diligences a raise before it commits need the same artefact: what was knowable while the decision was open, who said what on which document, and whether what was promised was delivered. That is what the record keeps, whoever turns out to be reading it.