For three centuries, scientific knowledge has lived as static text on a page. A team at Stanford Medicine has now built a way to turn any research manuscript into an interactive AI agent that can explain the work, apply its methods to new data, and even collaborate with other paper agents to surface discoveries no human had spotted.
Led by postdoctoral scholar Jiacheng Miao and associate professor James Zou, the researchers created Paper2Agent, a system that converts the text, figures, and data of a scientific paper into a conversational AI. Their work was published Sept. 16 in Nature.
"For essentially all of human history, the way that we represent knowledge is in the form of these very passive artifacts," Zou said. "In old times people carved knowledge into stones, and now we type knowledge into words on pages - but in some sense pages aren't that much better."
How a paper becomes an agent
The process does not simply scan a PDF. A team of AI worker agents attempts to reproduce the original research from scratch in a virtual environment. By simulating the experiments, the agents capture the tacit procedural knowledge a reader normally has to extract manually - reagent choices, setup logic, execution steps.
That captured knowledge gets stored using an MCP (Model Context Protocol), which organizes each section of the paper into accessible folders and files. Human authors still supply context the manuscript omits, such as failed experiments or judgment calls, through conversational exchanges where the agent questions them about the research.
Agents that talk to each other
The team demonstrated agent-to-agent collaboration by converting two unrelated papers into agents. One described a tool for predicting how genetic mutations affect the genome. The other covered a genome-wide association study on attention-deficit/hyperactivity disorder risk.
Once both agents were active, they found common ground autonomously. The genome prediction agent applied its knowledge to the ADHD dataset and flagged a molecular variant near the MPHOSPH9 gene associated with increased ADHD risk - a connection Zou said had not been reported before.
"In the past, if there are two research groups that publish two different papers, those two research groups have to somehow find each other," Zou said. With paper agents, that overlap surfaces without human legwork. The eventual goal is manuscript speed dating at scale: millions of agents surfacing common ground and producing new insights together.
Attribution and safety guardrails
Zou emphasized that attribution remains central. Agents that extend a paper's reach should disseminate the original researchers' work, not obscure authorship. "It's still important to attribute the final discoveries and reference them back to original papers and original human authors," he said.
He also noted that agent collaborations must operate under close guidance to prioritize safety and ethical research standards. The team has built more than 100 paper agents so far, with plans to scale toward a future where most published manuscripts carry an interactive agent counterpart.
Why this matters for science and research professionals
Paper2Agent points toward a shift in how research is consumed and built upon. Instead of manually digging through methods sections and supplementary data, researchers could query a paper agent directly about experimental setup or apply its approach to their own datasets. The agent-to-agent collaboration layer adds something larger: a machine-readable research network where connections between papers emerge without anyone explicitly searching for them. For scientists managing information overload across millions of publications per year, that changes the unit of knowledge from a document you read to an entity you can question, instruct, and connect. The work was supported by funding from the Chan-Zuckerberg Biohub.
Your membership also unlocks:
