A CSO That Never Touches Data: Stanford's Experiment

Key points

• The CSO agent only plans and assigns work, and it never touches the data.
• A reviewer agent checks every analysis, and the CSO sends work back based on its comments.
• The agents copy a human org chart, so an expert in each field can audit the matching agent.

Copying a drug company’s complicated early research organization

Picture the meeting room at a pharmaceutical or biotech company that is about to start developing a drug. The researchers are wondering which of many disease proteins to go after. On one side, a geneticist recommends candidates based on genome data. On the other, a cell biologist asks in which cells that gene actually works. A safety scientist warns whether a dose of the drug could cause deadly toxicity, and a pharmacologist argues about whether to make it a small molecule or an antibody. Pulling together evidence that is scattered across many teams and organizations into one decision takes a long time and a lot of money.

The lab of Professor James Zou at Stanford University (associate professor of biomedical data science) ran an experiment that moves this whole early drug-development process into an AI system. The project is called the Virtual Biotech. First author Harrison Zhang, an MD-PhD student, told the Stanford Daily that it is not a for-profit company but “a research project and an idea.” Stanford’s news office also described it as a biotech with “37,000 employees” and “no payroll,” but that is only a figure of speech. It is not a business that sells products, and it has no lab bench or real lab equipment. The research was first released as a preprint in February 2026, and then formally published in the journal Science on September 17, 2026. The related code was released under an MIT license so anyone can use it.

The Virtual Biotech covers the early, computer-based research stage of drug development. That means choosing which disease target to go after, checking whether touching that target would be safe, deciding the physical form of the drug, and learning lessons from past clinical trial failures and successes. The authors noticed that the key evidence in this process is split across several departments, which makes it hard to coordinate. In 2025, Professor Zou presented the Virtual Lab, in which AI agents collaborate like an academic research group. In an interview with the Stanford Daily, he explained that this new system goes beyond the level of a single lab and copies a far more complex drug R&D organization structure.

A CSO agent that never touches the data, and the specialist agents it works with

The Virtual Biotech built its organization from 11 agents. At the top is the chief scientific officer (CSO) agent. Inside the CSO’s office are two support agents: a chief of staff agent that writes the pre-briefing, and a science reviewer agent that scrutinizes the output. Below them, 8 scientist agents work in four departments.

The first department is target discovery and prioritization. It includes a statistical genetics agent that looks at genetic evidence such as genome-wide association studies, a functional genomics agent that handles CRISPR gene-editing screens and drug perturbation data, and a single-cell atlas agent that checks which genes are active in which cell types. The second department is target safety. A pathway and protein interaction agent, the single-cell atlas agent, and an FDA safety officer agent work together there. The FDA safety officer reviews drug side-effect reports, drug warnings, and gene knockout mouse data. The third department is drug modality selection. It consists of a target biologist agent that looks at protein structure and location in the cell, and a pharmacologist agent that reviews existing drugs and chemical probes. The last department is the clinical team, which has a clinical trialist agent and the FDA safety officer agent mentioned earlier. The single-cell atlas agent and the FDA safety officer agent use their expertise across two departments.

Virtual Biotech agent org chart

The agents run on Anthropic’s Claude Agent SDK. Each agent was given different instructions and tools. The research team built 10 data servers based on the Model Context Protocol (MCP) and connected more than 100 tools that search and run calculations over public databases such as Open Targets, ClinicalTrials.gov, and large single-cell atlases. The CSO, which sets the overall plan, and the scientist agents, which do the specialized analysis, use the Claude Sonnet 4.5 model. The chief of staff and reviewer agents, which check work and prepare briefings, use the smaller and faster Claude Haiku 4.5 model for efficiency and speed.

When a research question comes in, the system goes through six set steps. It starts with a person asking whether a certain protein is a good drug target for cancer treatment. In the second step, the CSO does not launch a costly large-scale analysis right away. It first interviews the person who asked, to make the research intent concrete. Meanwhile, the chief of staff agent writes a briefing document with background knowledge on the field, the latest trends, and the data the system holds. In the third step, the CSO breaks the problem into subtasks and assigns them to the right scientist agents. It decides whether to run the departments in parallel or in sequence, and it tightly controls what information passes between departments.

In the fourth step, the scientist agents query the data, write and run analysis code, and report the results. In the fifth step, the reviewer agent evaluates the analysis by three yardsticks. It checks whether the original question was answered precisely, whether the evidence is strong enough, and whether the analysis was thorough. In the last step, if the reviewer finds a gap or a claim without evidence, the CSO sends the work back to that agent with feedback that carries the reviewer’s points. Only when everything passes the criteria does the CSO write the final summary report.

The most interesting point in this flow is the limit placed on the CSO agent’s role. In the paper’s words, the CSO “never directly accesses data or performs analyses.” The CSO agent is not someone who runs analyses. It is a coordinator that knows what to ask and whom to ask, and thinks only about how to combine the answers from each field. It is like an orchestra conductor who coordinates the whole performance without playing the violin himself.

A lung cancer case and tens of thousands of copies show the collaboration at work

How this way of working operates in practice shows well in an analysis of the B7-H3 protein in lung cancer. The research team gave the system no access to outside web search, and the model’s pretraining knowledge stopped in January 2025. Once the study began, the genetics agent looked at the genetic evidence. The single-cell atlas agent noticed that B7-H3 appears to be most active in fibroblasts, the connective tissue cells around a tumor, where it seems to send signals that suppress immune cells.

The CSO that saw this report called in the reviewer agent. The reviewer pointed out a sharp gap. Single-cell data, in which cells are separated and analyzed one by one, loses the spatial information about where the cells were located in the tissue. The CSO immediately told the single-cell atlas agent to redo the work. The agent reanalyzed spatial transcriptomics data from 12 lung tumor samples and showed that fewer immune cells were in fact distributed near the spots where B7-H3 protein is concentrated.

Next, the clinical trialist agent analyzed public cancer cohort data and checked whether high B7-H3 expression goes with lower survival among cancer patients. Finally, the target biologist agent confirmed that this protein is exposed on the cell surface, and the pharmacologist agent concluded that it has no pocket structure for a small-molecule drug to bind to. The CSO gathered all the evidence and reached a final conclusion: develop a drug in the form of an antibody-drug conjugate, an antibody that carries a toxic drug and delivers it precisely to the target cells. This whole process took less than a day, and the API cost used was only $46.

The number “37,000 agents” the research team announced does not mean 37,000 different kinds of agents with different roles. It means that one specialist agent was copied tens of thousands of times and put to work in parallel. To find out which drugs succeed in clinical trials, the CSO agent proposed adding real outcome data to a public dataset of 55,984 clinical trials. The system copied the clinical trialist agent 37,075 times and assigned one agent to each Phase 2 and Phase 3 trial at the same time. Each agent checked ClinicalTrials.gov first, then went through academic papers and press releases in order, and carefully recorded the source of every value it entered. According to the authors, a huge job that would have taken about 1,839 hours with a single agent was finished in just 6 hours.

Through this analysis, the system found a pattern: drugs aimed at genes that are active selectively in only a few cell types were 48% more likely to reach the market in the end, and had a 32% lower rate of side effects. However, this clinical data is observational data, not the result of a controlled experiment, so it indicates a statistical correlation, not a causal relationship.

Where do people stand in this process? The paper describes it as “lightweight human interaction.” While the case analysis ran, people did only two things themselves. One was answering the CSO’s early questions to make the intent clear. The other was asking follow-up questions after each analysis round, requesting more analysis or sensitivity checks. The authors said, “Experimental ideation, design, and execution remained agent-driven.” Instead, people reviewed the results at the front and back of the system.

Every session leaves the agents’ reasoning and the tools they used visible on screen. The public code is designed so that each time it runs, it saves the code, the data, a list of claims, source information, and an audit report as files, so a researcher can rerun the analysis exactly. The authors had two human reviewers directly review 100 randomly chosen clinical trials and compare them with the agents’ results. The agreement was 89.7% for primary endpoint judgments, 83.9% for secondary endpoints, and 92.4% for adverse event rates. On 7,278 clinical trials that overlap with an existing dataset that people had built by hand, the agreement was 85.3%. Professor Zou told the Stanford Daily that a PhD-level expert redid several of the key findings by hand to confirm reproducibility.

Checking agreement on clinical trial data

There was also validation from the outside world. In August 2025, after the model’s training knowledge cutoff, the US Food and Drug Administration (FDA) gave breakthrough therapy designation to ifinatamab deruxtecan, a B7-H3-targeted antibody-drug conjugate developed by a drug company. The design direction the Virtual Biotech arrived at, without knowing the latest outside information, matched exactly with frontline research in the real pharmaceutical industry. Professor Zou called this “an independent, third-party validation.”

The differences that the division of labor and record-keeping create

What sets the Virtual Biotech fundamentally apart from an existing biotech is not simply fast computing. It is the way it splits the work and the way it keeps a record of the process.

First, analysis and judgment are clearly separated. The authors point out that in an existing organization, the person who decides the research direction also weighs the evidence personally. In the Virtual Biotech, by contrast, the CSO agent only makes plans, assigns work, and writes the summary report. Analysis is done only by specialist agents holding limited tools, and the validity of that analysis is checked by a completely separate reviewer agent. The authors of the paper point out that in a company, decisions can be shaped by “organizational dynamics, differing prior beliefs, or employee incentives.” But this is only the authors’ concern, and it has not been tested by a direct comparison with real pharmaceutical teams.

Second, the AI’s org chart becomes a map for human review. The authors explain that they designed the organization to enable “each agent to be audited by the corresponding human expert more easily.” A genetics researcher can review the code of the genetics agent, and a safety expert can watch the record of the FDA safety officer agent. In an existing pharmaceutical company, the grounds for many decisions scatter and disappear across slide decks and closed meeting rooms. This system, by contrast, saves the analysis code and data sources at every step. The authors wrote that this level of transparency is “rarely achievable” in a traditional drug research setting.

Third, organic handoffs of work between departments run as a default rule. In the B7-H3 case, when the reviewer pointed out the limits of the single-cell data, it led straight to a spatial data analysis, and that in turn led to another agent’s survival analysis. The authors point out that in an organization with a walled-off culture where departments are cut off from each other, this kind of smooth handoff “may not occur.” In the Virtual Biotech, the CSO agent carries out the handoff as a matter of course.

Fourth, the cost of huge repetitive cleanup work fell so far that checking everything became possible. In an interview with Stanford’s news office, Professor Zou said that if people had sorted tens of thousands of clinical trial records by hand, it would have taken years. This system assigned one agent to each trial and finished checking all of them overnight. The API cost for two in-depth case analyses came to only about $46 and $54. Professor Zou predicted to the Stanford Daily that “The cost will be less than $1 per agent.”

Lastly, people have stepped back to the outer edge of the actual execution. People focus on asking the first question, answering the CSO’s clarifying questions, giving feedback along the way, and auditing the final output. Complex idea generation, experiment design, and code execution are led by the agents. But the paper itself warns against leaving safety only to the reviewer agent. The reason is that within the system, “a single mistake can propagate through the virtual organization.” The authors stress that “it is still pertinent that human experts are involved in the workflow.”

Five lessons other industries can take

If the approach this study presents spreads widely, the R&D structure of the pharmaceutical and biotech industry could change a great deal. Until now, properly reviewing a single drug target has required experts from several departments to gather and pool their material. If a company can get a comprehensive early review report in a day for a few dozen dollars, a drug maker could screen far more target candidates up front within a limited budget. But as Singularity Hub pointed out, choosing candidates better by computer does not make real lab cell experiments, or clinical trials on patients that cost huge sums, any cheaper or shorter.

The landscape for data collection and specialist consulting services could also shake. The dataset used as the reference point in this study was material that earlier researchers had built by hand. The dataset from the full analysis of clinical trials this time, by contrast, was built and released by AI agents in just a few hours. If this trend settles in, the value of services that merely search, sort, and collect evidence could fall, and the added value could move to the interpretation and judgment that make decisions from what was collected.

Even outside drug development, any organization that handles complex knowledge work can learn something clear from the design of the Virtual Biotech. First, separate the coordinator agent that directs the whole from the analyst agents that touch the actual data. Only by stopping the coordinator from touching the data can you make it ask clear questions of the specialist agents and verify the output transparently at each stage. Second, to make it possible for people to supervise AI agents properly, it helps to split the agents by copying the actual human job org chart. That is because the person in charge of each field can intuitively audit the agent that matches their own job.

Third, make a reviewer agent that critiques the output a required step in the flow of work. You need a structure in which the reviewer steps in before the final answer and sends back work that falls short of the standard, in order to filter out hallucinations and leaps in logic. Fourth, instead of blindly trusting the whole job, draw a random sample and measure statistically how closely it agrees with human judgment. Only by holding a concrete error rate as a number, as in the 100-trial sample check, can you trust the system. Fifth, for large repetitive jobs, copy the agent tens of thousands of times and run the copies in parallel. That keeps one agent’s conversation context from getting contaminated and lets a long job finish in a single night.

Limits that have not been proven

The Virtual Biotech is an academic research project, after all. It is not a commercial company that serves real customers under deadline pressure. The success cases in the paper were picked by the researchers themselves, and figures such as processing time, cost, and agreement with humans are also mostly the authors’ own measurements.

The details in this article are based mainly on the preprint released in February 2026. For the formal Science paper published in September 2026, only the abstract could be read, so some specific numbers may have changed.

Also, the conclusions drawn from the clinical trial analysis are correlations from an observational study. The disease targets the system newly proposed have not yet been physically verified in a real biology lab. Professor Zou said the next step is physical testing in a real lab, and so far no independent critique from outside experts of this system has been found.

---

Sources

1. Stanford Medicine News Center — Virtual biotech company puts thousands of AI scientist agents to work on drug discovery — https://med.stanford.edu/news/all-news/2026/09/virtual-biotech-company.html — 2026-09-17
2. bioRxiv — The Virtual Biotech: A Multi-Agent AI Framework for Therapeutic Discovery and Development (Zhang, Eckmann, Miao, Mahon, Zou) — https://www.biorxiv.org/content/10.64898/2026.02.23.707551v1 — 2026-02-23
3. Science — The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development (abstract only) — https://www.science.org/doi/10.1126/science.aeg6779 — 2026-09-17
4. GitHub — TheVirtualBiotech README (Harrison Zhang) — https://github.com/harrisongzhang/TheVirtualBiotech — accessed 2026-10-22
5. The Stanford Daily — AI-powered biotech could shave years off of drug discovery — https://stanforddaily.com/?p=1299728 — 2026-10-14
6. Singularity Hub — Virtual Biotech Company Puts 37,000 AI Agents to Work on Drug Discovery (Edd Gent) — https://singularityhub.com/2026/09/18/virtual-biotech-company-puts-37000-ai-agents-to-work-on-drug-discovery/ — 2026-09-18