Stanford Designs New Drugs Using 37,000 AI Agents

Stanford University has built a 'virtual biotech company' that coordinates 37,000 AI agents and is applying it to actual drug discovery research. One of the drug candidate molecules designed through this effort has had its efficacy confirmed in independent validation experiments conducted by pharmaceutical giant Merck. James Zhou, Associate Professor of Biomedical Data Science at Stanford University, revealed this at the 2026 VB Transform conference.

An AI agent is an AI that, when given a goal, autonomously progresses with its work. Until now, the mainstream approach in development settings has been 'one agent per engineer,' with tools like Claude Code, the coding tool from Claude, being well-known examples. However, Zhou argues that an era in which this premise collapses is arriving. He contends that the next stage is not about creating 'a single smarter agent,' but rather about 'operating large numbers of agents as an organization.'

The research began with a small-scale 'virtual lab' composed of 5 to 8 agents. An AI serves as the professor (principal investigator), and AI students with different specializations hold regular meetings, mirroring the structure of an actual research lab. Furthermore, Zhou's team has also implemented a system where agents can attend a 'Stanford replica campus' and receive additional training (supervised fine-tuning) to deepen their expertise in their respective fields. This virtual lab succeeded in designing nanobody proteins targeting a novel coronavirus variant (small antibody-like molecules that bind to viruses), and the nanobodies designed by AI demonstrated higher virus-binding performance compared to conventional human-designed products, according to Zhou.

Building on this success, the team significantly expanded the scale, evolving from mimicking a single research team to a 'virtual biotech' that replicates the organizational structure of major biotech and pharmaceutical companies. In this system, an agent serving as Chief Scientific Officer (CSO) oversees the whole, with groups of agents positioned below in departments corresponding to drug target discovery, molecular design, safety evaluation, and clinical trials. Specialization continues to advance within each department; for example, in the target discovery department, there are separate agents handling genetic data and agents handling genomics and single-cell data.

Zhou's team also conducted comparative experiments where the same problem was solved by both 'large teams of specialized agents' and a 'single high-performance agent.' The results demonstrated that the cooperative system of multiple agents tends to be more robust even when errors accumulate and produces higher-quality solutions. This suggests there may be advantages to an architecture that distributes roles among many coordinated agents rather than concentrating computational resources on a single model.

What this research demonstrates is that the transition from a stage of using AI as a 'tool' to operating it as an 'organization' is becoming reality, at least in the field of scientific research. Furthermore, it can be positioned as a concrete model for designing environments where thousands of agents can collaborate by connecting existing databases to AI's command structure (orchestration layer). The fact that practical significance has been partially confirmed through independent validation by Merck suggests the possibility of impact beyond purely academic achievement. Going forward, how to safely control such large-scale multi-agent systems and how to ensure cost and reliability can be viewed as the next focal point of research.