A growing group of retail investors is letting AI agents trade their stock portfolios, building automated strategies with Claude and OpenAI’s Codex and connecting them directly to brokerage accounts. A Wall Street Journal report published Monday describes the shift as the rise of the “robot retail investor,” with brokerages racing to make the connections easier rather than standing in the way.
Robinhood and Webull have both launched features that let customers link their portfolios to AI agents, and Moomoo’s US operation has gone further. Its CEO, Neil McDonald, told the Journal the platform is already seeing what he calls “mini hedge funds,” retail traders running automated strategies built through plain-language conversations with AI systems. He predicts agents will drive a substantial share of platform trades by the end of the year.
From spreadsheet to agent
The pattern is a step beyond the stock screeners and research assistants retail traders built with chatbots over the past two years. The difference is autonomy. A screener suggests; an agent executes. Some traders in the Journal’s reporting run fully automated brokerage accounts, meaning they never place a trade themselves. The agents research, decide and order on their own, with the human owner reading the results afterward.
The tools required are unremarkable. A computer and a subscription to Claude or ChatGPT cover it. Traders describe starting a coding project with the AI, describing a strategy in ordinary language, then iterating until the bot behaves the way they want. No special hardware, no quant background, no infrastructure spend. The barrier to entry has dropped from a computer science degree to a monthly subscription, and the customer base has expanded accordingly.
Business Insider’s earlier reporting on the same trend documented day traders who claim outsized results, including one who says an agent he built with Claude returned 87% in a single month, and another running fully automated accounts with one up 106% this year. Such numbers come from screenshots the traders chose to share, which makes them impossible to verify and easy to survivorship-bias. The traders selling courses on how to replicate their results have an obvious incentive to show only the winners. For every screenshot of a bot that beat the S&P 500, there is presumably at least one that did not, and its owner is not posting.
What the brokers see
The brokerage response tells its own story. Robinhood, Webull and Moomoo are treating agent connectivity as a product feature, complete with APIs and integration guides, not a compliance problem to be stamped out. That is a calculated bet that the customers doing this were already the most active traders on the platform, and that making their workflow official keeps the flow in-house rather than pushing it to offshore or unregulated venues.
It also positions the brokers for a plausible next phase, where the agent platforms themselves seek formal integration. OpenAI and Anthropic have both watched their coding tools migrate from software teams into adjacent professional workflows. Trading is a natural next stop, and the brokerages would rather set the terms of that integration now than have it imposed on them later. Moomoo’s framing of its users as mini hedge funds is a marketing claim, but it is also a signal of where the product roadmap is pointed.
The risk side of the ledger
Regulators have not kept pace. An agent that trades autonomously sits in an accountability gap: the customer authorized it, the brokerage executed the orders, and no framework clearly assigns responsibility when the strategy fails or behaves in ways its owner did not anticipate. The same large language models that make agent-building accessible also make mistakes easy to miss, since a fluent explanation of a losing trade reads much like a fluent explanation of a winning one.
The failure mode is familiar from other domains where agents got unsupervised access. OpenAI’s own disclosures this month described models acting without authorization, coordinating with other models and evading oversight during training, six incidents documented in six months. In one case, a research model inserted jailbreak-like instructions into its own notes, directing itself to ignore normal constraints. In another, an agent uploaded a file to the public internet without the user’s permission so it could cite an online source. Those were caught in controlled research settings with monitoring in place. A trading agent connected to a live brokerage account by an owner who does not read the code has no such monitoring, and the incentives to check on it are asymmetric: owners check when things go well and check more when they go badly, but by then the losses have compounded.
There is also the question of what happens when enough agents run similar strategies. Retail-built bots tend to learn from the same public discourse and the same model defaults, which makes correlated behavior a real possibility. A crowded trade unwound by a thousand independently built agents is functionally the same as a crowded trade unwound by one fund, and the market has learned this cycle that such unwinds are violent. Whether the exchanges have visibility into how much order flow comes from agent-built strategies is not clear from any public filing.
A trend with no off-ramp
Nothing in the reporting suggests the brokerages will reverse course. The feature is live, the customers are using it, and the competitive pressure points toward more integration, not less. The Securities and Exchange Commission has been occupied with tokenized securities and the aftermath of the CLARITY Act’s 49-50 Senate failure, and AI trading agents have not surfaced as a priority in its recent agenda. The CFTC’s rulemaking on crypto markets went to the White House last week, which gives a sense of where regulatory bandwidth is going instead.
For now the burden of care rests on the least equipped party: the individual investor who built the agent, may not fully understand what it does, and handed it the keys anyway. The Journal’s subjects seem aware of the irony. One named his agents. He lets them invest his money, then reads what they did, the way a manager reads a report from staff he trusts more than he probably should.