37,000 AI agents ran a biotech company

The latest company to spin out of a Stanford Medicine lab has 37,000 employees — and none of them are human. There is no lab space, no lunch breaks, and no payroll. It is an AI-powered virtual biotech, the brainchild of associate professor of biomedical data science James Zou and graduate student Harrison Zhang.

“Our idea was to see how far we could push this. Could we create a biotech company that takes on everything from looking for drug targets all the way to designing clinical trials?” Zou said. “Could we have a fully agentic company that tackles the extremely complex challenges of drug discovery?”

The virtual company mirrors the org chart of a real one: a chief scientific officer agent leads research teams broken into specialized divisions that work in parallel — molecular target identification, safety analysis, clinical trial design. A paper describing the system, with Zhang as lead author and Zou as senior author, was published September 17 in Science.

About 90% of drug candidates entering clinical trials never reach approval, most often on efficacy or safety. Zou assigned a single agent to each clinical trial to pull out its safety and effectiveness data. The agents catalogued 55,984 trials in less than a week — work that would have taken human researchers years.

Then the agents did something Zou calls “quite interesting.” For trials with single-cell gene activity data, they built two scoring systems: one measuring how specifically a drug targeted a certain cell type, and one measuring bimodality — whether a gene's activity behaves like a light switch (on-off) or a dimmer.

Drugs aimed at switch-like, cell-type-specific genes were 40% more likely to advance from phase 1 to phase 2, 48% more likely to reach market, and came with 32% fewer adverse events. The pattern held across cancers and brain, heart, kidney, and lung conditions. Zou's theory: a target that behaves like an on-off switch and homes in on one cell type is easier — and safer — to control with a drug.

To test whether the virtual company could actually design a drug, the team pointed the agents at B7-H3, a protein lung cancer researchers have long eyed. Working only from data available before January 2025, the agents found B7-H3 highly expressed on fibroblasts near tumor cells, suppressing nearby immune cells — effectively cloaking the tumor. They designed an antibody-drug conjugate: a tag that homes in on B7-H3-carrying cells and delivers a toxic chemotherapy payload directly to them.

Months later, in August 2025, Daiichi Sankyo — developing with Merck — independently arrived at the same antibody-drug conjugate strategy against B7-H3. That therapy, ifinatamab deruxtecan, went on to receive FDA breakthrough therapy designation after a phase 2 trial in 187 patients. “This was really exciting as an independent, third-party validation consistent with the effects and the design proposed by the virtual biotech,” Zou said.

None of the agents' proposals has been tested in a laboratory or in patients. The evidence so far is computational and retrospective — the system found patterns in trials that already ran and designed drugs on paper. No independent group has reproduced the work, and the paper's authors are its only source for the figures.

Zou is not chasing the B7-H3 target further since another company is already shepherding it into clinics. But he says the virtual biotech has surfaced other candidate targets designed the same way. “Our next step is to bring the new findings from the virtual biotech into real labs and test how many hold up in the real world.” Humans, physical experimentation, and validation remain the conduit through which the AI has to prove itself.