A biotech company with 37,000 employees, and none of them are people

I read something yesterday that sounded like the setup for a novel, and then it turned out to be a paper in Science.

A team at Stanford Medicine built what they call a virtual biotech company. It runs on about 37,000 AI agents, sorted into groups that cover the whole drug development pipeline, with a virtual chief science officer at the top handing out the work. It comes from James Zou, an associate professor of biomedical data science there, and a graduate student, Harrison Zhang.

The part I liked is that they didn't just let it run and then admire the output. They tested it. The agents went through the results of roughly 56,000 completed drug trials and pulled out patterns for which candidates tend to work. Then the team boxed the system in: only information that existed before January 2025, and one question, what would you try for lung cancer. It came back with an antibody-drug conjugate aimed at a protein called B7-H3, also known as CD276. Months after that cutoff, a real pharmaceutical company landed on a very similar strategy on its own, and that approach later got FDA breakthrough therapy designation.

So the system reproduced a real decision it had no way of knowing about. That is a much better test than most AI science demos bother with.

I'm not going to pretend this means drug discovery is solved. Agents are cheap to spin up and easy to make look busy, and getting the right answer about the past is not the same as producing a new medicine. But an org chart with 37,000 non-human employees, doing work you can actually check, is a strange thing to have in 2026.

Source: https://med.stanford.edu/news/all-news/2026/09/virtual-biotech-company.html