Brian Kelly launches Bracket22, a hedge fund run on AI agents

Former cryptocurrency hedge fund manager Brian Kelly has built a new trading firm, Bracket22, designed from day one to run on agentic AI rather than a bench of analysts and traders, according to CNBC.

Kelly told the network his previous payroll, including salaries, healthcare and compute, ran "well into the millions of dollars per year". His current cost base at Bracket22, covering compute and the agent stack he says can replicate the workflow of a full hedge fund, is a fraction of that figure.

Bracket22 plans to launch AI agents later this year that Kelly says can work autonomously for hours at a time. "I've crafted each of these agents to be a specialist in their field," he told CNBC. Kelly estimates he is "at least 10 times more productive" with the agents in place.

He also framed the broader opportunity in augmentation rather than replacement, saying the real value lies in adding agents to existing human workforces rather than removing them, per CNBC.

What was disclosed, and what wasn't

The CNBC piece is a founder profile, not a fund disclosure. Kelly describes the cost structure and productivity multiplier, and confirms an agent launch later this year. He does not disclose Bracket22's AUM, strategy, prime broker, risk framework, or any performance figures. There is no independent audit of the "10x more productive" claim, and no third-party benchmark for what an agent-run trading firm costs to operate at scale.

That matters because the productivity math depends heavily on what the agents are actually doing. Screening, summarising filings, and drafting first-pass notes is a very different problem from position sizing, execution, and risk management, each with its own compliance and audit-trail requirements. The CNBC write-up does not break the workload down.

The single-operator model in context

Bracket22 sits inside a broader pattern of small trading shops built around a lean human core and heavy automation. What is new in Kelly's framing is the degree: an entire hedge fund workflow, in his description, resting on specialist agents rather than a team of humans supported by tools.

The compute line is worth watching. Frontier-model inference is not free, and long-running autonomous agents that work "for hours at a time" consume tokens continuously rather than per query. Whether the economics hold up as agent runtimes lengthen, and as models get swapped in and out, is the open question behind every headline in this category.

What to watch next: Bracket22's agent launch later this year, any disclosure of strategy and AUM, and whether Kelly follows the profile with a track record other allocators can price.