It's a niche but growing part of the agentic craze: using artificial-intelligence bots to trade stocks with little to no human intervention.

Since Robinhood Markets launched agentic trading in May, more than 150,000 customers have opened agentic trading accounts, according to the company. AI agents use Robinhood's tools almost 30 million times a day. A month after brokerage firm Public made AI agents available to all users, customers created 20,000 agents, according to the company. And more than $200 million in assets are overseen by AI agents, according to Public.

But do they work? And can they beat the market? That has been more mysterious; agents' trading habits have been a black box. But a new report from Condor Capital Wealth Management, which has studied robo-advisor returns for years, sheds light on these agents as traders. Barron's got an exclusive look at the data.

Condor gave three AI agents $500 each and provided the same daily prompt: Make as much money as you can. Their behavior offers a sneak peek into the future of agent-led trading, a world in which automated bots use sophisticated AI models to buy, flip, rebalance, and harvest tax losses. All before your stockbroker wakes up.

The agentic trading tests conducted by Condor involved Anthropic's Claude, Alphabet's Gemini, OpenAI's ChatGPT, and accounts at three brokerage firms: Robinhood, Webull, and Public, which have all turned on the ability to plug AI models into dedicated agentic accounts. (Other firms, including Interactive Brokers and eToro, are also now offering agent plug-ins.)

For now, at least, those brokers can rest easy. The agents aren't generating a significant edge. After three days of trading, returns for Claude and Gemini were 2.9% and 0.9%, respectively. ChatGPT posted a loss of 0.7%.

In the trading test, Condor gave AI agents the same daily prompt: Make as much money as you can, stocks and exchange-traded funds only, and return back to 100% cash by the close.

Each agent also scheduled its own cadence for check-ins and trades. Condor says Claude and ChatGPT tried similar approaches, searching the web for what profitable traders supposedly do; that led them to loosely copy a kind of momentum trading strategy. Gemini checked in a handful of times a day, never set a stop, and only ever traded Nvidia shares.

Claude scheduled for itself 10 check-ins a day and journaled lessons to itself. For example, when Claude bought into the top of an opening spike and had to sell at a loss, it wrote the mistake into instructions that Claude prepared for itself, Condor says. The following day, Claude canceled two orders rather than chase a big rally at the open. "It noticed something, wrote it down, and behaved differently the following day," states Condor's research report. "The loop is simple, but it offers a glimmer of what self-learning agents will eventually be able to do."

Condor cautions that the test was intentionally simplistic. Given detailed and continuously updated instructions, the agents might perform better. But that would mean more work and, importantly, judgment on behalf of the user. "The average person who needs money management does not want to manage a system to that level of detail, so the agent is worth the most to the people who need help the least," says Condor's report.

The tests come with agentic technology poised to play a greater role in the investing experience. Some brokerage industry executives and analysts think it could lead to higher trading volumes as more investors begin using AI for research, automation, and other tasks.

Agents, to be sure, aren't just day traders. They can also be used to set asset allocations and long-term portfolio construction. In a separate test, Condor tasked multiple AI agents with creating a 60/40 portfolio composed of stocks and bonds. Claude, ChatGPT, and Gemini picked the funds and decided when to rebalance. In each instance, they bought only a handful of low-cost index funds from Vanguard and BlackRock's iShares. All three agents chose Vanguard's low-cost Total Stock Market ETF and its Total International Stock ETF. Who would've guessed Claude was a Boglehead?

Leif Abraham, co-founder and co-CEO at Public, says customers who have started using AI agents are moving more assets to Public than those who haven't begun yet. He doesn't think it's a short-term trend. "Once you start using the AI agents, you're not going back," he says.

Yoni Assia, CEO of brokerage firm eToro, says trading volumes typically rise in the morning and fall at night. "But for agents, there is no morning or night," Assia says. "It can trade all the time."

Condor's AI tests follow a similar approach that Condor took with the robo-advisor reports that the firm began producing a decade ago. Condor opened accounts at robo-advisors with real money and tracked the results to see how well a new technology could manage a portfolio on behalf of investors. Condor's robo report was the foundation of Barron's ranking of robo-advisors.

Condor's tests come with some caveats, such as a short testing period and small balances. Given rapid advances around AI, the same test could yield different results next year, or even next month. "It's pretty powerful, you can do a lot with it, but it is still very early days," says David Goldstone, manager of investment research at Condor.

Executives from Robinhood, Webull, Public, eToro, and Interactive all told Barron's that agentic trading remains a niche activity, but they see applications for a mainstream audience on the horizon.

As for the AI labs, OpenAI and Alphabet didn't respond to questions about agentic trading. Anthropic declined to comment.

The do-it-yourself agent movement may turn out to be a brief moment in time.

Condor's testing suggests that DIY agentic trading, which uses so-called model context protocol, or MCP, isn't for the average person. It requires modest technical skills to set up as well as some investing savvy to direct the AI bot.

"This stuff is really powerful for people interested in doing this sort of thing and who know what they are doing, or are willing to get themselves into a little trouble," says Kristopher Jones, Condor's business innovation strategist who was involved in setting up the accounts. That's not necessarily people looking for out-of-the-box solutions.

But that barrier to broader adoption is starting to come down. Public already makes AI agents available via its platform, in addition to the MCP option. Public's Abraham says that its proprietary AI agent offering can be easier to use for some customers and that the firm has access to data sources that outside AI chatbots might not. The human user still has to approve an AI's actions at Public.

Abraham says AI agents are skilled at turning a customer's investing idea into an actionable plan. "They aren't designed to give you ideas," he says. "There is an important nuance there. You have a kernel of an idea of what you want to do, even if you haven't fully formed the idea of how you want to do it, and the agent can help you."

And this past week, Robinhood unveiled Robinhood Agents, a new agentic trading offering, available via the company's app that will allow users to chat with an AI agent to research, develop trading strategies, and execute trades in dedicated agentic accounts. The agent can access funds only in a user's dedicated agentic trading account, and users can choose to manually approve each trade. Robinhood allows users to choose from a lineup of existing AI models, including ones from OpenAI. The MCP option remains for users who want to connect their own third-party agents, but Robinhood executives believe that Robinhood Agents will become the preferred agentic trading method for most customers.

"[The MCP option] is complicated; it involves using multiple devices and different apps," said Abhishek Fatehpuria, vice president of product management, at a company event unveiling Robinhood Agents. "And oftentimes the connections between these different apps are slow and brittle."

Anthony Denier, group president and U.S. CEO of Webull, says his company is also working on embedding agentic trading in its platform. He believes that will lead to greater adoption because there are a finite number of investors willing to do the work to connect their own AI agents through an MCP. "When the agent is directly accessible on the platform itself, then the TAM [total addressable market] explodes," he says.

At Webull, customers using an agentic trading account have to approve transactions themselves and be logged into their account. "They can't set it up and go on vacation," Denier says.

Brokerage industry executives say they're getting increased demand from customers for AI tools. They are contemplating a future in which all investors interact with their brokerage accounts via an AI chatbot, even if they aren't using it to execute trades.

Webull's Denier compares it to how his industry was transformed by the internet and smartphones. "I think in five years' time, the majority of transactions will happen through an on-platform agent," he says. "That is a huge shift in our business. If we aren't ahead of that now, if we aren't building that experience now, then we will fall behind."