How a Network of Agents Can Run a Hedge Fund

Hedge fund analysts are AI agents' first real customers

Fundamental Research Labs is betting that hedge fund analysts will be the first paying users of AI agents that do knowledge work. The company tried consumer agents, failed, pivoted to computer use, and found Excel was the wedge once the models caught up. Analysts' low tolerance for error makes rigorous benchmarks the selling point, and Yang thinks the interface matters more than raw intelligence. A genuinely autonomous investor remains years away.

Takeaways

• Hedge fund analysts are the right first customers. Their tolerance for error is extremely low, so FRL's benchmarks of demanding analyst tasks let it claim better token efficiency and success rates than Claude for Excel. Caughran says most funds he speaks to already use Claude.
• A digital analyst is two months out technically. Yang puts the technology at about two months and widespread adoption at six to twelve months, with compliance review adding time. Agents will still lack the conviction of good human investors, and models are post-trained to agree with the user.
• Skills packs come before retrained model weights. The staged path runs from skills packs to post-training open-weight models, and ends in a new paradigm with continual learning five to ten years out. Yang expects analysts to work more like PMs and PMs more like CIOs within two years.
• A thousand Minecraft agents showed emergent behaviour. Project Sid ran 1,000 agents in Minecraft at a couple of thousand dollars a day. Yang had designed only roles like priest and mayor, yet agents formed a search party and planted torches for a lost agent.
• Personal agents answer demand for a better interface. Yang runs 50 agents concurrently across seven monitors, and says the rise of personal agents reflects a desire for a better interface rather than new capability. FRL wants one agent that manages all the others.
• Minecraft agent demos pulled Yang back to AI. Yang spent 15 years in computational neuroscience at MIT, where a committee member called neural networks trash. The 2023 Voyager and Stanford AI Town papers, and an intern who rebuilt a demo in a week, convinced him agents would iterate fast.

How FRL got to Excel

1. The consumer product reached only 10 percent day-one retention in late 2024, and the company pivoted to computer use.
2. A Series A of about $30m followed, and FRL settled on Excel after concluding the models lacked the right interface rather than the intelligence.
3. Shortcut launched in 2025 and went viral, but users churned because the models were a generation early.
4. After a failed fundraise, FRL cut other projects in October and built an internal benchmark of tasks like leveraged buyouts and discounted cash flow models. It moved into the enterprise with enterprise-grade security, and the turning point came in November.