Stanford Medicine and the Chan Zuckerberg Biohub have unveiled a fully autonomous “virtual laboratory” that uses teams of generative-AI agents to propose, debate and test biomedical hypotheses. Described in a peer-reviewed study published 29 July in Nature, the system is organized around an AI principal investigator that recruits specialised software ‘scientists’, including immunology, computational biology and machine-learning agents, and a built-in critic to identify flaws. In its first demonstration, the virtual lab was tasked with finding a better basis for a vaccine against new SARS-CoV-2 variants. Within days, the agents agreed on designing nanobodies—small antibody fragments that can be modelled rapidly—and produced structural blueprints using tools such as AlphaFold. Researchers at Chan Zuckerberg Biohub then synthesized the lead design and confirmed in vitro that it binds tightly to both recent and original COVID-19 spike proteins without off-target effects. Stanford associate professor James Zou, who led the work, said human researchers intervened in about 1 % of the automated process, largely for budget constraints. All agent discussions are logged for transparency, and the workflow can be rerun or audited. The team is now feeding experimental data back into the system to refine the nanobody and plans to apply the platform to other biomedical challenges. If broadly adopted, the approach could compress discovery cycles, lower costs for small laboratories and expand interdisciplinary research capacity beyond human working hours, the authors argued. The study highlights both the potential and the need for governance frameworks as AI agents assume greater autonomy in scientific research.