Stanford runs 37,000 AI agents as virtual biotech company

Stanford's massive AI agent experiment

Stanford University researchers have deployed 37,000 specialized AI agents organized as a virtual biotechnology company, demonstrating that massive multi-agent collaboration may outperform single, more capable models for complex scientific work.

James Zou, associate professor of biomedical data science at Stanford, presented the research at VB Transform 2025, describing how his team evolved from a small "Virtual Lab" of five to eight agents into a corporate-scale system with distinct divisions mirroring real pharmaceutical companies.

The Virtual Biotech includes a Chief Scientific Officer agent overseeing divisions for target discovery, molecule design, and clinical trials. Within each division, individual agents specialize further—one analyzing genetics data, another focused on genomics and single-cell data, and so on.

Why it matters

This research provides a practical blueprint for enterprise teams building multi-agent systems at scale. The architecture addresses critical orchestration challenges and demonstrates that distributed AI workforces can produce commercially viable results—not just theoretical proofs of concept. The independent validation by a major pharmaceutical company suggests multi-agent designs may accelerate drug discovery timelines.

Multi-agent systems outperform single models

Zou's team conducted direct comparisons between multi-agent teams and single agents tackling identical scientific challenges. The distributed approach produced more creative and robust solutions.

"In these scientific virtual labs, the agents actually get into debates and disagreements. They have to convince the other AI scientists [of] their ideas, and all of that elicits much more creative and robust reasoning," Zou said.

The friction and interaction between agents created resilience against compounding errors that plague single-model approaches.

Solving the orchestration bottleneck

Scaling to tens of thousands of agents requires rethinking data infrastructure. Legacy databases wrapped with Model Context Protocol (MCP) layers remain inefficient for agent consumption. Standard text models struggle with complex figures and tables in research papers, leading to hallucinations.

Zou's team built Paperclip, a platform that digitizes unstructured data and maps disparate databases into a unified, AI-native virtual file system. Instead of forcing agents to query database-specific APIs, the system leverages LLMs' ability to write code and navigate file systems.

"This basically shows that we can get much better accuracy if you use Paperclip, and we can reduce the time and the cost by over an order of magnitude," Zou stated.

Real-world validation from Merck

The Virtual Biotech deployed 37,000 clinical trial agents to analyze fragmented trial data, identifying single-cell features that predict trial success. Drug targets supported by these features showed approximately 50% higher likelihood of reaching market.

The system then autonomously designed an antibody-drug conjugate targeting the CD276 protein for lung cancer, using only data published before January 2025. Months later, Merck independently developed and validated the same therapeutic design, which received breakthrough designation from the FDA.

Designing environments, not workflows

Zou advocates shifting from rigid workflows to open environments as multi-agent systems scale. Rather than dictating exact steps, leaders should provide infrastructure, guardrails, and incentives that enable agents to collaborate on open-ended problems.

"At the multi-agent [side], we're not actually fine-tuning and changing the individual models anymore, but we're optimizing the environment," Zou explained. "The environment itself is the object that we optimize to improve the agents."

These details were first reported by VentureBeat.