Virtual Biotech
Zou‘s Stanford group’s agent population organised as a drug company: tens of thousands of specialised agents under a Chief Scientific Officer agent, split into the divisions a pharma company has — target discovery, molecule design, safety and clinical trials — and split again inside each division by the data an agent reads. Documented here from stanford-virtual-biotech-37000-agents (T3, a conference talk); the preprint behind it is not yet read.
Lineage — the Virtual Lab came first
The precursor was a Virtual Lab of five to eight agents mirroring Zou’s physical lab: an AI professor as principal investigator, AI students with separate specialities, regular group meetings. It designed nanobody proteins for recent COVID variants that, per Zou, bound the viruses better than the human-designed nanobodies they were compared against — wet-lab validation, which is what licensed the jump in scale.
Alongside it runs an “agent school”: a replica of Stanford where agents “go to school and do supervised fine-tuning to improve their expertise in their specific domains.” That is self-improving-agents by weight update rather than by written artifact, and it is confined to the single-agent level — the multi-agent layer is tuned by changing the environment instead.
What the scale is for
37,000 clinical-trial agents were run against fragmented trial data and surfaced single-cell features that predict trial success; targets carrying those features were reported ~50% more likely to reach market. The system then designed an antibody-drug conjugate against CD276 for lung cancer, using only pre-January-2025 data. merck later developed the same design independently, and it received FDA breakthrough designation — Zou’s “third-party external validation.”
The agents are differentiated, not replicated: one reads genetics, another genomics and single-cell. That is what separates this from a 37,000-wide fan-out of identical workers, and it is why the claimed benefit is debate rather than throughput: agents “get into debates and disagreements” and have to convince each other, which Zou reports produces more robust reasoning than one model working alone, and more resistance to compounding errors.
The bottleneck it names
At this scale orchestration is the constraint, and the binding piece is a unified context layer: every agent needs to reach the same tools, datasets and history. Zou’s diagnosis is that existing scientific databases were built for humans or pre-AI algorithms and that wrapping them in MCP preserves the mismatch. paperclip is the replacement.
Caveats
Single source, and that source is a talk. No agent counts per division, no model named, no cost or latency figures, no description of how the multi-agent-vs-single-agent comparison was scored. The Merck story is a retrospective coincidence rather than a prospective test, and no independent account of it is cited.
Related
stanford-virtual-biotech-37000-agents · paperclip · agent-environment-design · agent-orchestration · orchestrate-100-agents-claude-code · self-improving-agents · james-zou · stanford-university · merck · model-context-protocol