Stanford Builds a Virtual Drug Discovery Factory With 37,000 AI Agents

A biotech company with 37,000 employees would normally require humongous offices, extensive laboratories, clinical teams and computational scientists with an enormous operating budget. Stanford researchers hoverver have created something that looks remarkably similar on paper, except that none of its employees are human.

The Virtual Biotech Company, developed by researchers led by Stanford’s James Zou and described in Science on September 17, 2026, uses a hierarchy of AI agents to reproduce the organizational structure of a traditional drug-discovery company. A chief scientific officer agent receives a research question from a human, delegates work to specialized agents and combines their outputs into a larger scientific analysis. At its largest deployment, the system used more than 37,000 clinical-trialist agents to analyse nearly 56,000 clinical trials.

37,000+ AI clinical-trialist agents	55,984 clinical trials analysed	<1 week to process the trial dataset
40% higher Phase I→II progression associated with cell-type-specific targets	48% higher probability of reaching market in the analysis	32% fewer adverse events in the associated analysis

The important point is that Stanford researchers did not simply deploy thousands of copies of a chatbot and ask them the same question. In the same way an agent handles software development by breaking down a large task into smaller tasks by spinning multiple agents. Researchers created an organizational hierarchy in which different agents perform different scientific functions on smaller clearly defined tasks. The system can delegate tasks involving statistical genetics, functional genomics, disease biology, molecular targets, clinical data, pathways and interactions, then integrate the resulting evidence. The architecture resembles a company more than a conventional AI model.

From One AI Scientist to an Entire R&D Organization

The first major test involved clinical trials, which tradintinally is the slowest part of drug development. Instead of asking a single model to read tens of thousands of studies, the system assigned individual agents to specific trials. Those agents extracted structured information about outcomes and connected drug targets to genomic and single-cell information. The resulting dataset allowed the researchers to investigate whether particular biological characteristics correlated with clinical success.

The agents identified an association between drugs targeting genes with strong cell-type specificity and clinical outcomes. According to the study, drugs targeting cell-type-specific genes were 40% more likely to progress from Phase I to Phase II, 48% more likely to reach the market, and showed 32% fewer adverse events than the comparison group. The pattern appeared across multiple disease areas, including cancer, neurological, cardiovascular, kidney and lung diseases. These figures describe an association identified in the researchers’ analysis; they do not establish that cell-type specificity itself causes better clinical outcomes.

That distinction matters: Stanford’s Virtual Biotech drug discovery laboratory does not have clinical evidence that it has invented a new universal rule for successful drugs. What it demonstrates is that an agentic system can perform a large-scale evidence-integration task that would be extraordinarily difficult for a small human team to execute manually.

Bigger Challanges: AI Aided Cancer Therapy

The more interesting experiments follwed the clinal research trails. They investigated B7-H3, also known as CD276, as a potential lung cancer target. The system combined genetic, single-cell, spatial-transcriptomic and clinical evidence and focused on B7-H3 expression in fibroblasts surrounding tumors. The agents proposed that B7-H3-positive fibroblasts could contribute to immune suppression around the tumor and suggested an antibody-drug conjugate designed to target those cells and deliver a cytotoxic payload.

The researchers then performed an unusual form of retrospective validation. They restricted the Virtual Biotech’s information to data available before January 2025. Months later, in August 2025, an established pharmaceutical company independently developed the same broad B7-H3 antibody-drug-conjugate strategy. The external program later received FDA breakthrough therapy designation after showing efficacy in a human study. The Stanford team does not claim that its AI system discovered the clinical drug, but the independent convergence provided a real-world validation of the computational hypothesis.

That is an important distinction from many AI-drug-discovery announcements. The Virtual Biotech did not produce an approved medicine. It generated a therapeutic hypothesis and candidate strategy that subsequently converged with an independently developed real-world program.

Why Thousands of Agents?

The obvious question is why researchers needed 37,000 agents instead of a single AI agent. The answer is the same as it is with development: parallelization. If you have the computational power, why limit to solving/investigating to one agent and one problem. A human research team might divide a project into several workstreams. An AI system can divide a large dataset into tens of thousands of smaller tasks, assign each task to an agent and recombine the results.

That changes the economics of certain forms of scientific analysis.

Instead of: One Scientist → One Analysis → One Result the new workflow becomes: One Scientific Question → Thousands of Specialized Analyses → Integrated evidence → New hypothesis

The approach resembles a massive computational research organization. This is where the Virtual Biotech connects with the broader movement toward AI scientists. Labcritics has already covered Google DeepMind’s Co-Scientist and GPT-Rosalind, but Stanford’s approach introduces a different architectural idea: instead of trying to build one omniscient scientific model, create an organization of specialized agents.

The virtual biotech still needs a real laboratory

This is perhaps the most important limitation.

The Stanford system currently operates primarily in computational space. The researchers themselves emphasize that human scientists and physical experimentation remain necessary to determine whether its hypotheses survive contact with biology. The next step involves taking discoveries from the virtual company into real laboratories and measuring how many hold up experimentally.

That creates an obvious next stage:

This is where Stanford’s Virtual Biotech becomes part of a much larger transition in life sciences.

AI is no longer simply being asked to predict a protein structure or summarize papers. Researchers are beginning to construct systems that can organize scientific work, perform thousands of analyses, formulate hypotheses and determine what should be tested next.

The missing component is the physical laboratory to actually do the experiment. A bottlenect that AI and Pharma companies forsee and are actively setting up automated wet labs or dark labs to do experiments and feed the data loop.

The next question is much harder an the answer will be interesting: What happens when thousands of AI scientists are connected to thousands of automated experiments? That is where the virtual biotech could stop looking like a simulation of a pharmaceutical company and start looking like a new kind of scientific organization.

Related Labcritics coverage:
AI Biodesign Accelerator · GPT-Rosalind · LifeSciBench