How a 22-Year-Old Oxford Dropout Built the AI Agent Taking Over Wall Street

A portfolio manager who couldn’t stop laughing told me about Jonah Reuter for the first time. He wasn’t sure whether to be impressed or upset after losing a trade to a 22-year-old’s software. “The thing doesn’t sleep,” he remarked while stirring an espresso that he wasn’t consuming. “And it doesn’t blink.”

A few years ago, Reuter would have seemed like an unlikely founder. Halfway through his second year, he left Oxford, citing what he refers to as “an allocation problem” with the sly smile of someone who has said it too many times. His parents weren’t overjoyed. According to multiple accounts, his tutors were incensed. However, by the time the autumn term started without him, Reuter had already moved into an apartment above a Shoreditch kebab shop and had begun writing the initial framework for what would eventually become Anvil, an autonomous AI agent that is currently executing trades covertly inside two of the biggest investment banks globally.

It’s more difficult to determine what Anvil actually does than what the marketing portrays. In the traditional sense, it is not a quant model. A trader is not signaled by it. It scrapes Bloomberg terminals, watches news wires, reads earnings calls, parses regulatory filings, and then, and this is the part that worries people, argues with itself. The system, which Reuter loosely borrowed from research on recurrent-depth transformers—the same family of architectures behind a wave of bottom-up open-source projects this year—conducts internal discussions among sub-agents before issuing a single recommendation. As one developer recently posted on GitHub, it has the same weights, more loops, and deeper thinking.

Speaking with those who have used it, it seems that Anvil’s speed makes it unimpressive. Many things are quicker. It appears to hesitate at the appropriate times, which makes it impressive. Traders who have observed the system describe an odd sensation of working alongside something that hesitates, rethinks, and sometimes doesn’t do anything at all. Anvil turned down a trade that his desk was about to complete, according to a London PM. Twenty minutes later, unreleased news caused the underlying stock to plummet. He is still unable to explain how the model arrived before him.

Even Reuter is difficult to read. In person, he is slender, a little anxious, and wears the same grey sweatshirt in almost all of his photos. He discusses markets in the same way that some people discuss chess: admiringly, a little obsessively, and with a hint of disdain for those who don’t play well. It’s difficult to ignore the fact that he lacks the confidence of a founder pitching to investors when you watch him in the Shoreditch office, surrounded by whiteboards covered in equations and what appeared to be a half-eaten croissant. He exudes the concentration of someone who has already moved on to the next issue.

Wall Street has previously succumbed to younger, unfamiliar individuals. The old guard is still unsure of how to react to him. Speaking on condition of anonymity, a senior managing director at one bank informed me that his team has been told to examine Anvil’s outputs and draw lessons from them—basically, the business equivalent of conceding defeat. According to investors, Reuter’s agent may be among the first real instances of an AI system integrated into institutional trading’s regular operations. It remains to be seen if that holds true over the course of a complete market cycle. Since Anvil hasn’t really experienced a downturn, it’s still unclear how well it performs in one.

It has caused a conversation to change. It would have been a joke five years ago that a young Oxford student could create a trading agent that could effectively compete with quant desks at multibillion-dollar companies. The desk is now the punchline. The people who are watching this the most intently in London right now have the impression that something has tipped, not loudly or all at once, but enough that the old playbook now appears a little out of date. Reuter is twenty-two years old. He’s got time. Strangely, the banks might not.