37,000 Agents, One ADC: Stanford's Virtual Biotech and Merck's Independent Convergence

The dominant mental model for AI in drug discovery remains one engineer, one agent. At VB Transform 2026 on August 7, Stanford associate professor James Zou argued the next frontier is the opposite: tens of thousands of specialized agents organized like a biotech company — with a Chief Scientific Officer agent, divisional structure, and external validation from pharma.

From virtual lab to virtual biotech

Zou's team began with a five-to-eight-agent "Virtual Lab" mirroring his physical Stanford group — an AI principal investigator, specialized AI students, even an "agent school" for supervised fine-tuning by domain.

That setup designed nanobody proteins for recent COVID variants that, Zou reported, "worked much better than the previous human-designed nanobodies" in binding assays — wet-lab validation of multi-agent design.

The group then scaled to Virtual Biotech: tens of thousands of agents across target discovery, molecule design, safety, and clinical trials divisions.

Orchestration: Paperclip

Zou identified orchestration as the bottleneck at scale. Wrapping legacy databases in MCP servers, he argued, does not solve the underlying problem: "the interface and APIs are not suitable for agents."

The team's Paperclip platform maps disparate sources into a unified, AI-native virtual file system. Zou stated it reduces "time and the cost by over an order of magnitude" versus agents without AI-native scientific infrastructure.

37,000 clinical trial agents

Virtual Biotech spun up 37,000 "clinical trial agents" to synthesize fragmented trial data. Targets supported by agent-identified single-cell features were about 50% more likely to reach market than comparable drugs without them.

The system autonomously designed an ADC targeting CD276 for lung cancer using only pre-January 2025 data.

Merck as independent confirmation

Zou said Merck independently developed and validated the same therapeutic design, which received FDA breakthrough designation — "third-party external validation" of the virtual biotech output.

The framework is described in PMC (The Virtual Biotech: A Multi-Agent AI Framework for Therapeutic Discovery and Development), reporting annotation of 55,984 clinical trials.

Environments vs. workflows

Zou advocated designing environments, not workflows — infrastructure and guardrails over prescribed steps. Multi-agent teams produced better science through debate and disagreement, he reported.

Caveats

• Merck convergence is one data point; broader replication needed.
• Regulatory path still requires human-led IND packages.
• Compute cost at 37,000-agent scale is unspecified publicly.

### Sources

• VentureBeat — Stanford is running 37,000 AI agents as a virtual biotech — and one of its drug designs got independently confirmed by Merck (August 7, 2026)
• PMC — The Virtual Biotech: A Multi-Agent AI Framework for Therapeutic Discovery and Development (2026)
• James Zou — VB Transform 2026 presentation (August 7, 2026)