Stanford’s Virtual Biotech Deploys 37,000 AI Agents, Merck Confirms a Design

Stanford University has turned its biomedical data science lab into a virtual biotech powered by 37,000 autonomous AI agents. At the VB Transform 2026 conference, associate professor James Zou unveiled how the system mimics a corporate biotech, from target discovery to clinical‑trial simulation, and how one of its antibody‑drug designs was later validated by Merck.

Virtual Biotech Scales to 37,000 AI Agents

The architecture rests on a unified AI‑native file system called Paperclip, which converts PDFs, tables and figures into searchable code‑level objects. This sidesteps the “PDF‑in‑context” bottleneck that plagues traditional LLM pipelines.

“The agents actually argue, propose experiments and critique each other, which yields solutions that a single model would miss,” Zou told Reuters.

In a head‑to‑head test, the multi‑agent ensemble outperformed a monolithic model, delivering designs that were ↑ 50% more likely to succeed in downstream trials. The same platform generated an ADC targeting CD276 for lung cancer; months later Merck announced an independently derived version that earned FDA breakthrough status, a de‑facto third‑party confirmation.

Scaling the workforce required an orchestration layer that treats each agent as a micro‑service, with a Chief Scientific Officer agent coordinating divisions that mirror real‑world pharma units. The system also slashed computational expense by ↓ 10% compared with legacy pipelines, according to internal benchmarks cited by Bloomberg.

Zou argues the next wave of AI‑driven R&D will shift from rigid workflows to open environments where incentives, guardrails and shared data repositories drive collaboration, much like a self‑organising research campus.