At VB Transform 2026, James Zou, associate professor of biomedical data science at Stanford University, argued that the prevailing developer assumption — one engineer, one agent — is changing. He said the next frontier is not a single, more capable agent but architectures made of tens of thousands of specialized agents working together.
Zou’s presentation emphasized a practical problem for developers and product builders: how to orchestrate these massive multi‑agent systems. His team has published and demonstrated an approach for connecting legacy databases to AI orchestration layers and for building environments that enable thousands of agents to collaborate.
From a Virtual Lab to a corporate Virtual Biotech
The project began as a "Virtual Lab": a small group of five to eight agents designed to mirror Zou’s physical Stanford lab. The setup included an AI professor acting as the principal investigator and AI students with distinct specialties who held regular group meetings.
"We also created for the agents a replica of Stanford, an agent school, where the agents can actually go to the school and do supervised fine‑tuning to improve their expertise in their specific domains," Zou said.
The virtual lab successfully designed new nanobody proteins for recent COVID variants; Zou reported the AI‑designed nanobodies bound better to those viruses than previous human‑designed nanobodies.
After validating designs in wet‑lab experiments, the team expanded scope from modeling a single research team to modeling a large corporate structure. The resulting system, named Virtual Biotech, comprises tens of thousands of specialized AI agents overseen by a Chief Scientific Officer (CSO) agent, and it is organized into divisions such as target discovery, molecule design, and clinical trials.
"Working with the CSO agent are different divisions that mirror the divisions found in a human biotech or pharma company," Zou explained. Individual agents specialize further within divisions — for example, one agent may analyze genetics data while another focuses on genomics and single‑cell data.
Why many agents can outperform one
As foundation models become more capable, designers face a core architectural question: why distribute tasks across tens of thousands of specialized agents rather than concentrate compute in a single omniscient model?
Zou’s team ran head‑to‑head comparisons. The multi‑agent ecosystem created friction, debate and interaction that produced more creative and error‑resilient solutions than a single agent working alone.
"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 compared to if you have a single model trying to do the problem by itself from scratch," Zou said.
The orchestration bottleneck and Paperclip
When scaling to tens of thousands of agents, orchestration becomes the primary bottleneck. The system needs a unified context layer so agents can synthesize knowledge from tools, datasets, and historical records.
Many enterprise teams try to make legacy databases usable by wrapping them with middleware, but those systems were not designed for agent consumption. Dropping a PDF into an agent’s context is inefficient; language models struggle to interpret complex figures and tables, which can cause hallucinations.
To address this, Zou’s team built Paperclip. The platform leverages a core strength of modern LLMs — their ability to write code and navigate file systems. Instead of forcing agents to query brittle, database‑specific APIs, Paperclip digitizes unstructured data and maps disparate databases into a unified, AI‑native virtual file system.
That structure allows agents to access knowledge from millions of papers using standard file‑system operations.
"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 compared to if you use agents without these AI‑native scientific infrastructures," Zou said.
Real‑world test: 37,000 clinical trial agents and an ADC design
To test the architecture, Virtual Biotech launched 37,000 "clinical trial agents" to synthesize fragmented trial data. These agents identified single‑cell features that predict trial success — drug targets supported by these features were about 50% more likely to reach market than comparable targets without them.
The system then autonomously designed an antibody‑drug conjugate (ADC) targeting the CD276 protein for lung cancer. The design was completed autonomously and relied only on data published before January 2025.
Several months later, Zou said, pharmaceutical company Merck independently developed and validated the same therapeutic design; that program subsequently received breakthrough designation from the FDA. Zou described Merck’s independent validation as a third‑party confirmation of the virtual biotech agents’ therapeutic design.
Designing environments rather than rigid workflows
As multi‑agent systems scale, leaders must rethink how they manage digital workforces. Zou recommended moving from prescriptive workflows — which tell an agent exactly what steps to take — to designing open environments that provide infrastructure, guardrails and incentives while leaving collaboration open‑ended.
Optimization at scale, he argued, focuses on engineering the environment rather than fine‑tuning individual models. Individual agents can still be improved through reinforcement learning or supervised fine‑tuning in an agent school, but the performance of a massive multi‑agent system depends on tuning the parameters that govern collaboration.
"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."
Why this matters
Zou’s work suggests an industry shift: the future of AI application development may emphasize building ecosystems of many specialized agents and AI‑native data infrastructures rather than pursuing ever larger single models. Success will depend not only on model capability but on how data and collaboration environments are organized to let agents work together effectively.
