"Houston, pull up the strategy charts Desmond generated over the weekend."

Brian Kelly spoke a single command to his computer, and several charts instantly appeared on the screen.

In the past, that request would have traveled across several desks: quantitative researchers compiling test results, analysts drafting charts, and then materials being routed to the fund manager. Today, Desmond is an AI agent who worked through the weekend, and Houston is also an agent handling the transfer of information.

Kelly is the founder of the hedge fund Bracket22. The team also includes Steffi and Doocy — one monitors technical signals, while the other is dedicated to attacking investment logic. These four names make up an entire investment research team, and the only human in the group is Kelly.

This is not his first fund.

In early 2025, Kelly shut down his previous cryptocurrency hedge fund. That company employed only seven or eight people but was responsible for covering multiple countries and time zones. Between salaries, bonuses, health insurance, and office rent, Kelly estimates the annual operational cost was roughly $5 million.

A few months later, he began intensively testing AI tools. When rebuilding Bracket22, Kelly did not rehire his former team. Instead, he dismantled the structure of a fund and reassembled it piece by piece using AI agents.

Building an Agent for Every Role

Many companies introduce AI as an assistant to their existing team. Kelly took the opposite approach: he first broke the team down into job functions, then assigned an agent to each role.

Desmond handles quantitative strategy, while Steffi focuses on technical analysis. Once research is complete, Doocy steps in. Acting as a red team, its job is not to polish materials but to find flaws and attempt to tear down the entire investment thesis. Houston sits at a higher level, calling on other agents, collecting results, and presenting differing opinions to Kelly.

In this organizational chart, research, rebuttal, coordination, and decision-making are deliberately kept separate. Kelly also keeps his specialist agents isolated from one another, wanting each to form its own conclusions without influencing the others. Ultimately, he relies on his own experience to make the final call.

Capturing Institutional Memory

Kelly replicated more than just a few job positions. He also built a "corporate brain," connecting research materials, historical trades, and emails into an interlinked node system. In his analogy, these nodes act like neurons: when a new signal appears, the system can locate similar past research; when reviewing a trade, it can retrace the original reasoning along the recorded chain.

This is the hardest asset to see in any investment research firm.

A traditional fund's memory is scattered across hard drives, email inboxes, meeting minutes, and employee minds. When someone leaves, context that never made it into written reports disappears with them. Kelly wants to transfer that experience away from individuals and into a system that agents can continuously access. As long as records keep accumulating, this one-person company can maintain an institutional memory that grows over time.

Of course, isolating agents doesn't guarantee truly independent opinions. If they use similar models and data sources, they could start from different roles yet converge on the same blind spots. Doocy can simulate the opposing side at an investment committee, but it won't bear the consequences of a dissenting opinion like a real partner would.

That is why Kelly always retains his own judgment for final decisions.

Costs Cut From $5 Million to $40,000

The first thing to change was the cost structure.

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

What was eliminated isn't just seven or eight salaries, either.

The cryptocurrency market operates 24/7. Previously, to keep the fund running around the clock, Kelly had to deploy staff across different time zones, along with benefits, office space, bonuses, and management overhead.

Now, agents don't need shift changes, and they don't shut down at 3 a.m. Desmond can test strategies all weekend, and when Kelly arrives at work, Houston presents the results.

Kelly estimates his personal productivity has improved at least tenfold. He even speculates that a 100-person firm equipped with AI could achieve the output of what once required 1,000 people. Research coverage is becoming decoupled from headcount.

Previously, a fund that wanted to cover more asset classes or run more strategy variations had to keep hiring analysts and engineers. Now it can simply add more agents. The fixed labor costs that once had to be committed in advance are transforming into model and compute expenses that scale with actual usage.

A New Competitive Landscape

This shift will first change the barrier to entry for small asset managers and proprietary trading firms. A fund manager with a strategy and capital no longer needs to assemble a full team before gaining research coverage that was previously only affordable for institutions.

Meanwhile, the business of financial AI companies will also evolve: clients no longer just want a chatbot that summarizes research reports. They want a complete digital workforce—something that finds signals, something that plays devil's advocate, something that coordinates, and a "brain" that remembers every transaction made in the past.

Funds are beginning to resemble software companies: once the system is built, adding new capabilities no longer requires paying per head.

Saving Money Isn't the Same as Creating Alpha

Costs are down, but what about investment performance? Kelly hasn't provided the most critical numbers, and that is exactly where the controversy lies.

For any fund, charts, reports, and 24-hour availability are just intermediate products. The final deliverable is returns. Bracket22 has disclosed its personnel structure, costs, and productivity, but it has not disclosed its rate of return, maximum drawdown, Sharpe ratio, win rate, or performance relative to a benchmark.

People can confirm that Kelly made a trading firm cheaper, but they cannot confirm that these agents are better at making money than his old team.

The investment industry frequently confuses two different report cards. One records "how much work was done": how many reports were read, how many markets were covered, how many strategies were tested. The other records "what results the work produced": how much additional profit was generated while bearing the same level of risk.

AI can easily make the first report card look impressive. The second one is much harder to improve.

Lower costs certainly have value. If investment performance remains unchanged, eliminating millions of dollars in annual operating expenses would itself improve Kelly's net returns. Some small strategies that weren't worth testing in the past may re-enter consideration because the cost of trial and error has dropped. But saving money is not the same as creating Alpha, and more research doesn't automatically mean more valid signals.

On the contrary, running too many strategies can increase the odds of stumbling upon spurious patterns in historical data. As more institutions adopt similar models, data sets, and frameworks, agents could also push similar strategies into the same trades even faster. What took weeks to become crowded in the human era could take only hours with machines.

So Bracket22 has passed the operational test, but it hasn't yet passed the investment test. Kelly has replicated the assembly line that produces investment judgments. Whether Alpha can be replicated is a question the market still has to answer.

Can Large Institutions Replicate This Model?

It's unlikely.

Kelly can compress his firm down to one person primarily because he is trading his own money. With no external LPs, he doesn't have to explain to clients whether a drawdown was caused by a model, nor does he need to convince anyone to entrust their capital to an agent named Houston.

Self-funded capital gives Bracket22 an enormous margin for experimentation.

Once client money comes into play, the equation changes. A trade doesn't just need to be directionally correct; it also needs a record of who approved it, what the rationale was, and who bears responsibility when something goes wrong. Fiduciary duties, compliance reviews, permission segregation, and audit requirements don't disappear just because researchers become agents. Agents can generate opinions, but they can't sign contracts on behalf of a firm.

Most Wall Street institutions are currently pursuing human-AI collaboration: Morgan Stanley routes some work to AI; Goldman Sachs places systems like Devin and Claude into teams where people and AI work together; BNY even gives its "digital employees" login credentials, email addresses, and permissions—yet still assigns them human managers.

The caution of large institutions isn't simply slowness. What they need to preserve isn't just people; it's a chain of accountability where responsibility can be traced when things go wrong and where checks and balances exist when disagreements arise. For a one-person firm, organizational redundancy is a cost. For a firm managing client capital, redundancy is sometimes an insurance policy.

Talent pipelines face the same dilemma. A junior analyst is an expense today, but years later that person could become someone who can spot anomalies and challenge models. If breaking down financial statements, testing assumptions, and drafting initial research are all delegated to agents, where do young employees learn through mistakes, and how do they develop judgment? Inside Goldman Sachs, there are concerns that excessive reliance on AI could weaken the reasoning abilities of junior bankers. Skimping on today's training costs may also deplete tomorrow's senior talent pool.

So the Bracket22 case functions more like a stress test. It demonstrates a new way for funds to operate: an investment firm can be assembled like software, specialized labor can be replicated by agents, and organizational size no longer determines research coverage.

At the same time, it pushes the fund manager's role into an even more concentrated position—machines expand cognitive bandwidth, but the human remains responsible for objectives and results.