Stanford Builds 37,000-Agent AI System to Simulate Full Drug Development Pipeline

Stanford Medicine researchers published a study in Science on September 17, 2026, detailing a virtual biotech company composed of tens of thousands of AI agents covering the entire drug-development process. Led by Zhang and senior author Zou, the system identified candidate success signals and proposed an antibody-drug conjugate targeting B7-H3 using data predating January 2025. Notably, a pharmaceutical company independently reached a similar strategy months later, which subsequently received FDA breakthrough therapy designation — serving as an external validation of the system's outputs. The research highlights key design principles such as role specialization, shared memory, and human oversight over irreversible decisions. Experts note the findings carry broad implications for deep-tech startups, emphasizing the need for structured agent governance and rigorous audit trails over simply scaling up AI agents.