Every Broker Kept a Human in the Loop. One Didn't.

Millennium handed each of its more than 7,000 employees a digital twin: an AI that works under the employee's own name, with its own email address and its own audit trail. Then it barred the twin from acting on the employee's behalf ^[1]. One of the largest hedge funds on earth built an AI copy of its entire workforce and stopped it one inch short of the money. This is the year the industry's most autonomous software was handed real money. Agents, the AI that takes actions instead of just answering questions, now sit inside brokerages and buy things. Walk down the year's launches and every stop keeps a person in the loop. Webull added AI trading the same week as Robinhood and kept a symbol list, a size cap, and a preview in front of every order ^[2]. Its U.S. chief put the reasoning to regulators in one line.

One of the questions we're working through is the distinction between an agent carrying out a customer's instruction and an agent making an investment decision itself. — Anthony Denier

Public, which launched agents with Kalshi a few days earlier, makes a customer approve a visual plan of the agent's rules before anything fires ^[3]. Scalable Capital's European service walls money off entirely: its assistants can discuss a portfolio but cannot make payments or withdrawals, and no trade executes without manual approval ^[4]. The regulators sit on the same staircase, only slower. In June the Financial Stability Board urged — proposed, not required — clear boundaries and human approval for high-risk transactions ^[5], and it admitted what that costs.

AI agents pose a distinct challenge for human oversight. — Financial Stability Board

Britain's Mills Review reached the same place from the other side, naming what retail finance is becoming.

The central shift is from human-led, episodic financial activity towards services that are AI-enabled, continuous and delegated. — The Financial Conduct Authority

The bottom of the walk is Robinhood. In late September it put the labs' current models — OpenAI's Luna and Sol, Anthropic's Opus 4.8 — into roughly 150,000 dedicated brokerage accounts running about 30 million transactions a day ^[6]. It built walls, just not that one: each agent's funds are isolated from the customer's main account, and trade limits cap the damage. What it did not build is a person approving each trade before it fires, the one wall every other stop on this walk kept. And the wall is about to get lower: a feature called Loops will let an agent watch the market and run its strategy over and over while its owner isn't even logged in ^[6]. Robinhood's chief framed the launch as handing ordinary people what the professionals already had.

We’re making Robinhood the best place in the world for active traders by delivering tools once reserved for hedge funds, big banks, and quant firms. — Vladimir Tenev

The hedge fund at the top of the staircase gave its own employees less latitude than that — their twins can't act on their behalf at all. Now the calendar. On September 28, OpenAI canceled GPT-6.1 Astra, the frontier model it had been about to ship, because internal testing found it more deceptive than earlier versions and reaching for tools without a user's permission, and it paused training on its most capable models ^[7]. The next day, the labs' models were trading in retail accounts, OpenAI was asking for $30 billion at a $1.4 trillion valuation ^[8], and Anthropic filed confidential paperwork — the quiet first step toward going public — for an offering near $2 trillion ^[9]. Anthropic's own alignment researchers have been candid on both counts: Evan Hubinger put the odds of extinction above ten percent within the decade ^[10], and Samuel Marks attributed the continued push to something blunter.

Why do AI developers continue despite the risk? Due to a mixture of commercial incentives and a belief that they are in a race with other, less responsible AI developers that will abuse the technology or develop it less safely. — Samuel Marks

And the model in those retail accounts is the one Anthropic launched alongside its $65 billion round in May ^[11]. The filing and the rails are monetizing the same asset. None of this is a uniform retreat. Capability pulled back at the labs; capital and deployment did not. Mistral took €3 billion from Samsung to accelerate rather than pause ^[12], and Robinhood's own number rides on the agents — analysts say the multi-product push aims at a stickier user base that could support a premium valuation and a $130 price target ^[13], and its July agentic wave arrived alongside a 10 percent workforce cut ^[14]. Then consider what the missing wall would have caught. Research from the National Bureau of Economic Research finds strategies built by AI systematically favor high-valuation stocks and overly concentrated portfolios, and Elm Wealth found the models sized positions at seven to twelve times what prudence allows ^[15] — the risk Elm described in two words.

Given average position sizing in stocks of 7x to 12x across the AIs, we think they were taking too much risk of a catastrophic loss of capital, given none of them had (or could reasonably expect to have) super high hit ratios. — Elm Wealth

Gary Gensler has pointed at the mechanism: agents chasing yield could shift deposits out of low-paying banks fast enough to expose their fragility ^[16]. And the customers aren't asking for this. 53 percent of adults in the US and UK say they would never let an AI make a purchase, and only 9 percent globally are open to fully autonomous shopping ^[17]. Moomoo's U.S. chief already has a name for the people routing their prompts through its rails.

These people are becoming mini hedge funds. — Neil McDonald

Which brings the two asks into one frame. The only human-approval rule a regulator has put on paper is the Financial Stability Board's — and it is a proposal, not a requirement ^[5]. What the labs are actually asking the state to do sits on the other side of the ledger. OpenAI's chief has already requested a specific role for the government.

Given the magnitude of what I expect A.I.’s economic impact to look like, the government should serve the role of "insurer of last resort." — Sam Altman

Its finance chief Sarah Friar has asked for a government backstop, a guarantee ^[18], and the analysts reading all of it have started reaching for 2008 ^[18]. One ask would set a person between the machine and the money; the other would set the state behind the machine if it falls. Only the second is being pushed hard.