After shutting down the "traditional research-oriented" approach, he used AI Agent to create an "AI-era hedge fund."

"Houston, bring up the strategy chart that Desmond ran over the weekend."

Brian Kelly spoke a sentence into the computer, and several charts immediately appeared on the screen.

In the past, this instruction would have passed through at least several desks: quantitative researchers would compile test results, analysts would draw charts, and then the materials would be delivered to fund managers. Now, Desmond, who works overtime on weekends, is the Agent, and so is Houston, who is responsible for relaying the information.

Kelly is the founder of the hedge fund Bracket22. The company also includes Steffi and Doocy, one who analyzes technical signals and the other who specializes in attacking investment logic. These four names form an investment research team, with Kelly being the only human.

This is not his first fund.

In early 2025, Kelly shut down the cryptocurrency hedge fund she had previously run. The company had only seven or eight employees, but it covered multiple countries and time zones. Kelly estimated that the annual costs, including salaries, bonuses, health insurance, and office rent, were approximately $5 million.

A few months later, he began intensively testing AI. When rebuilding Bracket22, Kelly did not bring back the old team, but instead dismantled a fund and reassembled it piece by piece using agents.

Strictly speaking, Bracket22 is not a traditional hedge fund that manages funds for external LPs. It only trades Kelly's own money, covering cryptocurrencies, stocks, and commodities.

But this also makes it a purely experimental question: what's left of an investment firm if there are no clients, colleagues, and offices?

Replace each job with an Agent

Many companies introduce AI to add an assistant to their existing teams. Kelly does the opposite: first, he breaks down the team into roles, and then assigns an agent to each role.

Desmond was in charge of quantitative strategy, and Steffi was in charge of technical analysis. After the research was completed, Doocy came into play. It played the role of the red team, not responsible for making the materials more attractive, but only for finding loopholes and trying to overturn the entire investment logic. Houston stood at a higher level, calling on other agents to collect results and then sending different opinions to Kelly.

In this organizational chart, research, rebuttal, coordination, and decision-making are deliberately separated. Kelly also isolated the professional agents, hoping they would make their own judgments first and not influence each other. Finally, he would make the final decision based on his own experience.

Kelly replicated more than just a few positions. He also built a "corporate brain," connecting research data, historical transactions, and emails into interconnected nodes. In his analogy, these nodes are like neurons: when a new signal is detected, the system can find similar past research; when reviewing a transaction, it can trace back to the initial judgment by following the records.

This is the most difficult asset for investment research institutions to see.

Traditional funds have their memories scattered across hard drives, emails, meeting minutes, and employees' minds. When someone leaves, much of the context not documented in reports also disappears. Kelly wants to extract this experience from individuals and place it into a system that agents can continuously access. As the records accumulate, this one-person company can also possess institutional memory.

Of course, separating agents doesn't equate to obtaining truly independent opinions. If they use similar models and data, they might enter the same blind spot from different perspectives. Doocy can simulate the opposing side on an investment committee, but unlike a real partner, it won't bear the consequences for a single dissenting opinion.

Therefore, Kelly always reserved her own judgment when making the final decision.

Costs reduced from $5 million to $40,000

The first thing to change is the cost.

In an interview with CNBC, Kelly stated that Bracket22's total annual expenditure on agents and computing power is approximately $30,000 to $40,000. Compared to the $5 million cost of his old fund, the new cost represents only 0.6% to 0.8% of the original, a reduction of over 99%.

It's not just seven or eight salaries that are being saved.

The cryptocurrency market operates 24/7. In the past, to keep the fund running 24/7, Kelly had to deploy staff in different time zones and bear the costs of benefits, office space, bonuses, and administration.

Now, the agent doesn't need to hand over shifts or shut down at 3 a.m. Desmond can test strategies all weekend, and then Houston can retrieve the results when Kelly comes to work.

Kelly estimates that his productivity has increased at least tenfold. He even envisions a 100-person organization with AI potentially achieving the output of a 1,000-person organization in the past, as research coverage is becoming decoupled from the number of employees.

In the past, if a fund wanted to cover another asset class or run another strategy, it would typically need to hire more analysts and engineers. Now, it can increase the number of agents first. The fixed human resource costs that used to be borne upfront are now being converted into model and computing power costs that vary with usage.

This will first change the barriers to entry for small asset management and proprietary trading firms. A fund manager with strategy and capital will no longer need to assemble a full team to obtain research coverage that was previously only affordable to institutions.

At the same time, the business of financial AI companies will also change: customers will not just want a chat box that summarizes research reports, but a whole digital team - someone to find signals, someone to contradict, someone to coordinate, and a "brain" that can remember all past transactions.

The fund started to resemble a software company: once the system was built, adding new capabilities no longer required per capita payment.

But saving money doesn't equal creating Alpha.

Costs were saved, but what about the return on investment? Kelly didn't provide the most crucial figure, which is also the most controversial part of this case.

For funds, charts, reports, and 24/7 online availability are merely intermediate products; the final product is the return. Bracket22 disclosed personnel, costs, and productivity, but not returns, maximum drawdown, Sharpe ratio, win rate, or performance relative to the benchmark.

Therefore, one can confirm that Kelly made a trading firm cheaper, but cannot confirm that these agents are more profitable than the previous team.

These are two report cards that are often confused in the investment industry. One records "how much work was done": how many research reports were read, how many markets were covered, and how many strategies were tested; the other records "what results were produced by the work": how much more money was made after taking the same risks.

AI can easily write a beautiful report card for the first one. The second one is much more difficult.

Lower costs are certainly valuable. If investment performance remains unchanged, saving millions of dollars in operating expenses annually will improve Kelly's net income; some previously untested strategies may also re-emerge due to reduced trial-and-error costs. However, saving money does not equate to creating alpha, and increased research volume does not necessarily translate to increased effective signals.

Conversely, running too many trades can make a strategy more susceptible to stumbling upon random patterns in historical data. As more institutions use similar models, data, and frameworks, agents may also cause similar strategies to crowd into the same trade more quickly. What takes humans weeks to create, machines might accomplish in hours.

Therefore, Bracket22 has passed the operational test but not the investment test . Kelly has replicated the production investment judgment pipeline; whether it can replicate Alpha remains to be seen, depending on the market's performance.

Can large organizations replicate this model?

It's difficult.

Kelly can compress the company into one person, primarily because he trades his own money. Without external LPs, he doesn't need to explain to clients whether a drawdown was caused by the model, nor does he need to convince others to entrust their funds to an agent named Houston.

Bracket22's own capital gave it a lot of room for trial and error.

Once client funds are involved, the situation changes. A transaction not only needs to be on the right track, but also needs to leave a record of who approved it, the basis for it, and who will bear the responsibility if problems arise. Fiduciary duty, compliance review, segregation of authority, and audit requirements do not disappear simply because a researcher becomes an agent. An agent can generate opinions, but cannot assume responsibility for the institution.

Currently, most of the AI explorations by major Wall Street institutions involve human-machine collaboration: Morgan Stanley has outsourced some of its work to AI; Goldman Sachs has placed systems such as Devin and Claude into teams where humans and AI work together; BNY has even given "digital employees" login accounts, email addresses, and permissions, but still assigns them human managers.

Large organizations are cautious not entirely because they are slow to act. They need to retain not only people, but also a chain of checks and balances to ensure accountability in case of problems and to prevent disagreements. For a one-person company, organizational redundancy is a cost; for organizations managing client funds, redundancy is sometimes insurance.

The same issue applies to talent development. Junior analysts are an expense today, but in a few years they might become people capable of identifying anomalies and questioning models. If breaking down financial statements, making hypotheses, and writing first drafts of research are all delegated to agents, where will young employees learn from their mistakes and form their judgments? Some within Goldman Sachs worry that over-reliance on AI will weaken the reasoning abilities of young bankers. Saving on training costs today might also mean mortgaging tomorrow's top talent.

Therefore, the Bracket22 example is more like an extreme test, demonstrating a new way of fund operation: investment research institutions can be assembled like software, professional division of labor can be replicated by agents, and organizational size no longer determines research coverage.

At the same time, it also pushes the role of fund managers to a more centralized position—machines are responsible for expanding cognitive bandwidth, while humans are responsible for goals and returns.

This article comes from the WeChat official account " AI Native Lab " , which continuously analyzes real-world AI implementation cases and shares enterprise AI practices and methodologies.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.