The Hedge Fund Run by Four AI Agents
Brian Kelly says four AI agents replaced a multimillion-dollar research team and made him ten times more productive—but one human still controls the money.
A hedge fund manager has shown CNBC a staff that never sleeps, never asks for a bonus and costs less per year than one junior Wall Street salary.
Brian Kelly’s Bracket22 is built around four AI agents. Houston acts as mission control. Doocey attacks investment ideas as a red-team critic. Steffi—named after tennis champion Steffi Graf—handles technical analysis. Desmond develops and tests quantitative strategies.
Kelly says the entire system, including computing, costs approximately $30,000 to $40,000 a year. His previous operation reportedly spent about $5 million on a team spread across New York, Hong Kong and California. He estimates that the agents make him at least ten times more productive.
The appeal is obvious in cryptocurrency markets, which trade around the clock. A human team needs shifts, handovers and management. Agents can continue reading, coding, testing and challenging one another at 3 a.m., then present a synthesized view when Kelly wakes.
But the viral description—“every employee is AI”—needs one crucial qualification. Kelly remains the human who decides whether money moves. He calls himself the “meat in the chair.” The bots generate analysis; the legal and financial responsibility stays with him.
That distinction is not cosmetic. Models can hallucinate data, overfit strategies or reinforce one another’s errors. Giving each agent a personality and specialty creates useful workflow separation, but it does not guarantee independence. If all agents rely on similar models or datasets, apparent debate may hide a shared blind spot.
The cost comparison also mixes two different organizations. A former multimillion-dollar fund carried compliance, operations, investor relations and personnel expenses. A lean experimental firm using cloud models may not perform all the same functions. Token spending is not the full cost of running regulated investment money at scale.
The system remains important because it previews how professional work may change. The likely near-term model is not a fully autonomous fund with no humans. It is one portfolio manager, one or two accountable professionals and a swarm of agents performing research that once required a larger junior team.
That could democratize sophisticated analysis for small firms. It could also eliminate entry-level roles through which future senior investors traditionally learned the business. If the apprenticeship ladder disappears, Wall Street may gain short-term efficiency while weakening its future human expertise.
Productivity is not performance
A research system can produce ten times more charts, code and summaries without generating ten times more investment insight. Financial markets punish crowded strategies and reward information that other participants have not already priced. If thousands of funds use similar frontier models, their recommendations may converge and disappear as an advantage.
Agent speed can even amplify instability. A false data point passed from one bot to another may acquire the appearance of consensus. Red-team agent Doocey is meant to challenge ideas, but its skepticism is only as strong as its instructions, data access and ability to recognize an unfamiliar failure mode.
Governance becomes central when outside capital is involved. Regulators and investors need audit trails showing which data informed a recommendation, what model version was used and why the human approved the trade. “The AI decided” cannot become a way to avoid fiduciary responsibility.
Cybersecurity creates another exposure. Agents with access to market feeds, code repositories and execution systems can be manipulated through poisoned data or prompt injection. Separating research from trading permissions is therefore as important as improving the models.
Kelly’s experiment is most persuasive as a demonstration of organizational leverage, not proof that humans are obsolete. Senior judgment, investor trust and legal accountability remain human functions. The disruption may fall first on analysts whose work consists of gathering, cleaning and presenting information—unless firms use agents to expand their capabilities rather than simply remove them.
What to watch next
Will Bracket22 publish audited returns, drawdowns and error rates? How does it prevent agents from inventing sources or leaking sensitive strategies? Four bots can produce more work at lower cost. The decisive question is whether they produce better risk-adjusted decisions when markets behave in ways their training data never anticipated.