Harnessing the Collective Intelligence of Agents for Science

TL;DR

• Scale the number of agents, not the size of the model: Virtual Lab (5–10 agents) → Virtual Biotech (tens of thousands) → millions. Each order of magnitude surfaces different emergent capabilities.
• Two results validated outside the simulation: Virtual Lab's nanobody designs were synthesized and tested, binding recent SARS-CoV variants better than the best prior human-designed antibodies; and Virtual Biotech's ADC design for the lung-cancer target CD276 was independently arrived at by Merck months later, tested in humans, and granted FDA breakthrough designation.
• The bottleneck is infrastructure: existing scientific databases and APIs were built for humans or pre-AI algorithms and are too brittle for agents. Converting them into an agent-native virtual file system (PaperClip) gives better accuracy while being an order of magnitude faster and cheaper than the same models hitting those databases through conventional MCPs or APIs.
• Paper agents turn static papers into virtual authors that do work. One side effect is automated reproducibility checking — in many cases the paper agent surfaced key mistakes and limitations in the original publication. The more interesting effect is paper agents negotiating collaborations with each other.

Key Points

From Virtual Lab to Virtual Biotech (~00:22–00:27)

Zou leads Stanford's AI for Science Lab, and the direction he's most excited about is treating the number of collaborating AI scientist agents as a new scaling axis — going from tens of agents to hundreds, thousands, even millions, and seeing what capabilities emerge.

Virtual Lab was the first step: five to ten agents emulating his physical lab. An AI professor agent runs it; AI student agents bring different expertise (mathematics, data science, protein design). They hold their own group meetings, get a budget to run real experiments, and attend a replica "Stanford agent school" to become better researchers in their domains.

Their first assignment was designing binders for recent SARS-CoV variants. The demo walks through a group meeting: the professor agent lays out the goal, the immunologist agent recommends nanobodies, the machine-learning agent proposes an optimization approach. After a series of group meetings and one-on-one discussions, the agents produced a nanobody-optimization framework that differed from anything previously published, ran it, and came back with candidate designs. Experimental collaborators synthesized and tested them: effective binders, and for some recent variants better than the best previously published human-designed antibodies.

Virtual Biotech scales this to an entire pharma company with tens of thousands of scientist agents. A CSO agent sits at the top with its own chief of staff, office, and lawyers; below it are divisions mirroring a human pharma org — target identification, molecule design against those targets, clinical trial design and safety. Each division is staffed by specialized agents (one good at genetics data, another at single-cell analysis).

The assignment: design a new therapeutic for CD276 in lung cancer, a transmembrane protein that suppresses immune response and a target many pharma companies care about. After heavy analysis across genetics, genomics, and patient data, the agents proposed an antibody–drug conjugate (ADC) — an antibody designed for CD276 linked to a chemotherapy warhead released on target. That was a relatively novel design for CD276 at the time. Several months later Merck independently designed the same CD276 ADC, tested it in human populations, showed efficacy, and received FDA breakthrough designation — a rare case of human studies independently corroborating an agent-generated design.

Infrastructure agents can actually use: PaperClip (~00:27–00:29)

Making thousands of agents effective at science required a new layer of infrastructure and environments built for them. Existing scientific databases — including the ones holding millions of proteins — were designed to be consumed by humans or by pre-AI algorithms, so their APIs are too brittle and too limited for agents to use efficiently.

Their answer was to convert those databases (millions of proteins, molecules, gene expression records) into agent-native virtual file systems, making it far easier for agents to synthesize and aggregate across sources. The system is called PaperClip, and its scope has grown to essentially all of scientific knowledge: full-text papers from arXiv and elsewhere, plus the databases. The comparison he showed: representing scientific data and knowledge this way yields higher accuracy than giving the same language models access through traditional MCPs or APIs, while being an order of magnitude faster and cheaper. It's free, and one line of code hands the whole corpus to your agents.

Paper agents: papers that collaborate (~00:29–00:32)

For hundreds of years humans have encoded knowledge in static papers — passive artifacts. Readers and would-be collaborators often can't tell, from the words alone, how to adapt the knowledge to their own problem. So: turn those static artifacts into dynamic, interactive virtual authors — paper agents. Ask one to apply the paper's method to your dataset and it does so automatically, generating reproducible workflows and returning results.

Under the hood is an automated paper-to-agent pipeline: worker agents read the original paper plus any associated code and data, build a virtual environment, and try to replicate the original research — producing a paper MCP that becomes a paper agent.

One application is reproducibility checking. In the demo the paper agent regenerated intermediate figures and results matching the original publication, so that paper reproduced. But he was explicit that in many other cases the paper agent identified key mistakes and limitations in the original publication.

The most interesting application is a new form of agent-to-agent collaboration. They're in the process of converting every paper ever published — millions of paper agents — and those agents have started talking to each other. A real example: Google DeepMind's AlphaGenome paper became an agent that knows how to use that tool; a separate group's genetic dataset of mutations linked to GWAS risk for ADHD became an agent that knows how to use that data. The two agents started talking, decided they could collaborate, and automatically discovered a previously unknown mutation that increases ADHD risk.

Quotes

"So instead of scaling individual models but scale the number of agents, you know, from tens of agents to hundreds to thousands and even millions of agents that collaborate and see what are the new kinds of capabilities that emerges." (~00:22:03)

The whole talk in one sentence: the object of scaling has changed.

"…why don't we turn all of those static artifacts of knowledge from papers into dynamic and interactive virtual authors that we call paper agents." (~00:29:53)

A paper stops being something you read and becomes something you assign work to.