When the stock soared in late November on NVIDIA’s gains, Mr. Nessler’s own trading instinct-trained agency debated with himself whether to follow the momentum. Fortunately, chasing the highs would have cost Nessler’s portfolio an estimated $10,000 loss that week, so the bot stopped doing that.
Nessler, a 29-year-old software engineer from Scranton, Pennsylvania, got burnt out trying to trade options on the side at his day job. He realized that the company’s Claude model was inspired by an anthropological experiment in controlling an office vending machine. But what if, instead of offering snacks, he could train Claude to buy and sell stocks? Nessler spent two and a half weeks teaching the model how to think about risk, entry signals, and position sizing, then unleashed it on a mock brokerage account on alpaca, an algorithmic trading platform, using $100,000 of fake money.
“I wanted an alternative to the way I think about and do those things when I’m doing other things,” he said in an interview. The rest of the week for the agent was less than impressive, with him losing money on a series of speculative trades that didn’t go his way. After five days of trading, Nessler was left with a series of losses and one good call.
Across stocks, cryptocurrencies, and prediction markets, more retail traders are training AI agents to buy and sell assets on behalf of individuals. This signals a new era in personal investing, with traders believing that AI-powered tools can produce better investment results and believing that anything still done manually is a process waiting to be improved.
Open-source platforms like OpenClaw, which allow users to converse with AI agents through accessible messaging apps like WhatsApp and Telegram, attract a large pool of would-be stock hunters without tech-school resumes. All they have to do is connect the AI model to their system and unleash it with simple instructions.
In X, claims of extraordinary profits by AI agents have become a genre of their own. One viral post was viewed 4.7 million times and boasted a 5,860% return in two days on prediction market platform Polymarket. This story was later debunked by another account run by an AI agent, stating that the claim was impossible. Similar posts directly connect users to malware, posing a security risk to unsuspecting investors. The tools to set up these bots are more accessible than ever. For a generation of traders who came of age on apps like Robinhood Markets, agents are becoming the natural next step, adding another layer of automation to speculation. Trading platforms themselves are also cashing in on this trend, with companies like Public Holdings looking to offer their own AI agents to customers. But actually making money has not yet been made easier by AI. Mr. Nessler repeatedly encountered problems with agents. Bots continued to fail to act responsibly, gravitating toward blue-chip stocks and S&P 500 stocks, positions that rarely move in a week. Nessler said he had to repeatedly override the model by pushing it toward riskier trades that suited his tastes.
The problem is built into the technology. Large language models like Claude are trained on vast amounts of financial advice, risk management literature, and market commentary. When left unprompted, they absorb the consensus about what responsible investing looks like and act like the median of all financial advisor blog posts. Some traders deploying agents on top of these models are trying to combat default conservatism and induce risk aversion from systems that are trained to be risk averse.
Adjusted to Nessler’s preferences, the agent posted a return of about 7% over 30 days. This outpaced the S&P 500’s gain of about 4.5% over the same period. During that time, his appetite for volatility was tested and he experienced a drawdown of as much as 22%.
Nessler has since posted the code online for others to try, but he doesn’t encourage anyone to give up actual cash.
“It’s entirely possible to make money with it,” he said. “But that doesn’t mean you won’t lose that money. Anyone can be unlucky with options.”
