Jeff Dean unveils Discovery Loop, an agent-driven research startup

Jeff Dean has founded Discovery Loop after leaving Google, with the goal of scaling up AI self-improvement and automating the full research workflow. He introduced the company publicly for the first time at a Stanford HAI event, where he laid out both the technical direction and early progress. In his telling, the bet is that AI should not only solve problems but also design its own experiments, validate the results, and keep learning from earlier work. Rather than large teams running experiments one at a time, he imagines a small number of domain experts overseeing tens of thousands of AI Agents that explore different research directions in parallel.

To make that possible, Discovery Loop has built its own Agent Harness, an infrastructure layer for coordinating large numbers of agents working together. According to the company, roughly 20 automated improvement loops are running today, with a target of boosting domain-expert productivity by 20 to 50 times.

On the team side, Dean said researchers had joined from model projects such as Gemini, DeepSeek, and Kimi, along with contributors from the Go programming language project. In thefirst three weeks after founding, the company held about 400 interviews. He added that around 10 percent of the team have received the ACM Prize in Computing, and another roughly 10 percent are medalists from the International Olympiad in Informatics, mixing seasoned large-scale AI research engineers with younger research talent. The same event also featured a conversation with Fei-Fei Li, in which she questioned the idea of AI recursive self-improvement. As of this report, the company is at the stage of sharing its infrastructure, loop designs, and long-term vision.