Paper2Agent turns research papers into interactive AI agents

A team led by Stanford computer scientist James Zou has built Paper2Agent, an open-source framework that converts an academic paper, together with its accompanying code, data, and other supplementary material, into an interactive AI agent. Rather than just answering questions about a paper the way a tool like Google's Gemini Notebook (formerly NotebookLM) does, a Paper2Agent agent can actually run the methods the paper describes, and potentially combine them with tools built from other papers. Zou and colleagues described the system in Nature on 16 September. They tested it across statistics, econometrics, and astrophysics, but built their proof-of-concept demonstrations in computational biology, where turning a published method into something usable is often especially painful.

The main demonstration used AlphaGenome, a deep-learning model that predicts how DNA mutations affect gene regulation. Fed AlphaGenome's documentation and code, Paper2Agent produced 22 working tools in about 45 minutes with no human intervention, running on a personal laptop for less than $15 in computing costs. The tools covered different slices of AlphaGenome's functionality: predicting how a DNA change affects gene activity, comparing those effects across tissues, or analyzing several variants at once. A testing agent validated all 22 tools by running them against reference results, diagnosing and attempting to fix any that failed, with up to six attempts per tool before it would be dropped. The validated toolkit was packaged as a Model Context Protocol server and connected to Claude Code, producing a plain-English AlphaGenome assistant that, according to the researchers' own analysis, outperformed both standard Claude working from the raw AlphaGenome codebase and a specialist AI co-scientist tool called Biomni.

The team then agentified two more papers, one on how inherited DNA variants linked to autoimmune disease disrupt cell function, another on systematically silencing every expressed gene in immune cells, and linked them to the AlphaGenome agent. Asked to investigate the genetic basis of psoriasis, the three agents together identified the little-studied gene GPR137 as a likely causal factor and proposed 10 ways to test that inference. A human researcher chose one of the ten; the resulting analysis found that silencing GPR137 produced changes in gene activity strikingly similar to those caused by the psoriasis-linked variant in immune cells.

Other researchers see broader promise in the idea. "The idea of making papers more dynamic and executable through an agentic interface is quite compelling," says Dongping Chen, a computer scientist at the University of Maryland in College Park. Olivier Elemento, a computational biologist who directs the Englander Institute for Precision Medicine at Weill Cornell Medicine and who peer-reviewed the study for Nature, calls it "a real advance in terms of how we think about the publication process, with AI at the center and in a way that makes publications more interactive." Artur Skowroński, head of application development at the Polish software company VirtusLab, wrote in a blog post that Paper2Agent could also help bring papers to life in classrooms, letting students experiment with published methods instead of only reading about them.

The approach has real limits: of 100 computational biology papers the team tried, 26 could not be converted into agents, usually because of incomplete code, missing documentation, or other software that would not run. Zou treats that as a feature rather than a bug, since a failed conversion exposes gaps or errors in a paper's own materials: "Agentification itself is a useful certificate that says, 'This work is relatively complete and well documented.'" He says human scientists will keep the final say, but expects agents to become part of how papers are published, alongside data and code availability statements a paper might one day carry an "agent availability" statement, a virtual corresponding author on call around the clock. The day after the Nature paper appeared, Zou's team unveiled Virtual Biotech, a multi-agent platform modeled on a drug-development company, described in Science alongside a Paper2Agent-generated version of that paper. They also fed the Paper2Agent manuscript into Paper2Agent itself, producing an agent that now lives at paper2agent.ai, a paper about turning papers into agents that has turned itself into one.

Key facts

• Stanford's James Zou and colleagues built Paper2Agent, an open-source framework that turns a paper plus its code and data into a chat-based agent able to run the paper's own methods; described in Nature on 16 September.
• Applied to the AlphaGenome DNA-mutation model, it produced 22 working tools in about 45 minutes for under $15 in compute, entirely without human intervention.
• The resulting AlphaGenome agent reportedly outperformed both plain Claude given the same codebase and a specialist tool called Biomni, per the researchers' own analysis.
• Linking three paper-derived agents together, the team used them to flag the gene GPR137 as a likely factor in psoriasis and proposed 10 ways to test the idea; a human-chosen follow-up analysis backed it.
• Of 100 computational biology papers tried, 26 could not be converted into agents, usually due to incomplete code or missing documentation.

Why it matters

Zou frames Paper2Agent as a change in what a scientific paper is: "Knowledge should not be static records. It really should be dynamic and interactive, and this has many benefits, including making knowledge more reproducible but also enabling all sorts of new kinds of discovery." The problem it targets is familiar to anyone who has tried to reuse someone else's research code: an undocumented repo, broken dependencies, and a half-finished readme standing between a published result and actually trying it on new data.

Who it affects

The demonstrations were built for computational biology researchers specifically, though the team also tested the framework on statistics, econometrics, and astrophysics papers. Artur Skowroński of VirtusLab points to a second audience: educators, who could let students run published methods directly instead of only reading about them. The broader target is the academic publishing process itself, which is why the work went through Nature's peer review, with Weill Cornell's Olivier Elemento serving as a reviewer.

How to use it

Paper2Agent is open source. Given a paper along with its codebase, data, or other supplementary material, it automatically extracts the paper's core workflows and builds a tested, runnable toolkit, packaged as a Model Context Protocol server that connects to Claude Code or another compatible chat-based assistant. The researchers' own paper about Paper2Agent has itself been run through the tool, and the resulting agent is live at paper2agent.ai.

How solid is it

The work is peer-reviewed and published in Nature, and each generated tool goes through an automated testing agent that runs it against reference results, tries to fix failures up to six times, and drops a tool it cannot fix. The AlphaGenome agent's edge over plain Claude and over Biomni comes from the researchers' own analysis rather than an independent evaluation, and the flagship demonstrations were concentrated in computational biology even though the underlying method was tried on other fields too.

Risks and caveats

Conversion is not guaranteed: 26 of the 100 computational biology papers the team tried failed to become agents, usually because of incomplete code, missing documentation, or dependencies that would not run, meaning the approach works only as well as the underlying paper's materials. Zou says human scientists will keep the final say over agent-driven findings, underscoring that the psoriasis result relied on a person choosing which of the ten proposed validation approaches to run.

“Knowledge should not be static records. It really should be dynamic and interactive, and this has many benefits, including making knowledge more reproducible but also enabling all sorts of new kinds of discovery.”

— James Zou, Stanford University