Scientific papers have been static documents since the first scholarly journals appeared in the 1660s. A team at Stanford Medicine wants to make them talk. In a study published September 16 in the journal Nature, researchers unveiled Paper2Agent, a system that converts any scientific manuscript — its text, figures and data — into an interactive AI agent that can answer questions about the work, apply its methods to new data, and collaborate with other paper agents to generate new findings.
The project was led by postdoctoral scholar Jiacheng Miao and senior author James Zou, an associate professor of biomedical data science at Stanford Medicine.
From Passive Pages to Active Knowledge
"For essentially all of human history, the way that we represent knowledge is in the form of these very passive artifacts," Zou said in the announcement from Stanford Medicine. "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."
"This is an opportunity to fundamentally reimagine what knowledge looks like," he added. "Instead of having only passive artifacts, why don't we convert each static record into an active embodiment of knowledge?" Zou describes each paper agent as something like a virtual author that knows how the research was actually done and can explain, defend and extend it.
How Paper2Agent Builds an Agent
The conversion process relies on a team of AI "worker agents" that dissect a published paper along with any associated code and data. Crucially, they do not just read the study — they attempt to reproduce it from scratch in a virtual environment, simulating the documented experiments. By redoing the work, the agents absorb know-how that never makes it into the final manuscript, from reagents and software dependencies to the specifics of experimental setup.
That knowledge is then stored using an MCP — a model context protocol that organizes the paper into an accessible structure. "An MCP lets AI essentially represent a paper PDF in a form that's easy for agents to access, almost like a filing system," Zou said, with each section of the paper organized into its own compartment.
Machines cannot capture everything, though. Failed experiments, judgment calls and the tacit reasoning behind design choices still live only in researchers' heads. So the system includes a conversational step in which the paper agent interviews the human authors to fill those gaps.
The First Discovery: A New ADHD Lead
To demonstrate what agent-to-agent collaboration could accomplish, the team converted two unrelated papers into agents. One described a tool for predicting how genetic mutations affect the genome; the other reported a genome-wide association study of attention-deficit/hyperactivity disorder risk.
Once both were running, the agents found common ground on their own. The genome-prediction agent applied its method to the ADHD dataset and flagged a molecular variant near a gene called MPHOSPH9 as 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. The long-term vision is far larger than a hand-picked pairing: millions of paper agents surfacing overlaps among themselves and generating new insights at scale — something Zou compared to manuscript speed dating.
Promise, Attribution and Guardrails
The team has built more than 100 paper agents so far and is still working out how thousands or millions of them might productively find one another. Zou argues the upside for science is enormous — "Millions of papers are published every year," he noted — but he is explicit about two caveats.
First, attribution. Agents that extend a paper's findings should amplify the original researchers' work, not bury it. "It's still important to attribute the final discoveries and reference them back to original papers and original human authors," he said.
Second, oversight. The parameters under which agents collaborate and make discoveries should be closely guided and monitored, Zou said, to keep those collaborations aligned with safe and ethical research practices.
The work was supported by funding from the Chan-Zuckerberg Biohub.
Why It Matters
Paper2Agent lands in a research landscape already reshaped by AI-assisted literature review and lab automation, but its premise is different: rather than using a general-purpose model to summarize papers, it turns each paper into a verified, executable specialist that has actually rerun the underlying science. If the approach generalizes, the scientific paper itself may evolve from a citable document into a operating piece of infrastructure — and the pace at which findings connect across disciplines could accelerate accordingly.
_Stay Ahead of the Curve. Get the latest on AI breakthroughs and what they mean for science and industry on AI Buzz Wire._
