Paper2Agent Converts Research Papers Into Executable AI Agents

Scientists Can Now Query and Execute Published Research

A new open-source framework called Paper2Agent is transforming how researchers interact with published scientific work. Instead of reading a paper and manually implementing its methods, scientists can now convert publications into interactive AI agents that execute analyses, answer questions, and even collaborate with other paper-agents.

Developed by Stanford computer scientist James Zou and colleagues, the system takes a research paper along with its code and data, automatically extracts core workflows, and generates tested, runnable tools. The result is an AI agent that doesn't just discuss the paper's contents—it can actually perform the methods described within it.

Why it matters

Reproducibility remains a persistent challenge in computational research, where broken dependencies, incomplete documentation, and abandoned code repositories often make published methods unusable. Paper2Agent addresses this by creating a quality gate: papers that successfully convert to agents demonstrate completeness and proper documentation. For research teams, this could accelerate discovery by making methods immediately applicable to new datasets without weeks of implementation work.

From Static Text to Dynamic Tools

The researchers demonstrated Paper2Agent's capabilities using AlphaGenome, a deep-learning model that predicts how DNA mutations affect gene regulation. According to the study published in Nature on September 16, the system generated 22 validated tools from the AlphaGenome documentation in approximately 45 minutes on a personal laptop, costing less than $15 in computing resources.

The framework includes an automated testing agent that validates each tool against reference results, attempting up to six fixes if tests fail. Successfully validated tools are packaged into a Model Context Protocol server and connected to AI assistants like Claude Code, creating a user-facing agent that responds to plain-English queries.

Multi-Agent Research Collaboration

Zou's team extended the concept by linking multiple paper-agents together. When three agents—covering AlphaGenome, autoimmune disease variants, and gene silencing effects—were prompted to investigate psoriasis genetics, they collectively identified GPR137 as a likely causal gene and proposed validation approaches. A subsequent analysis confirmed that silencing this gene produced gene activity changes similar to those caused by psoriasis-linked variants.

"Knowledge should not be static records," Zou explains. "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."

Limitations and Quality Control

The system isn't universal. Of 100 computational biology papers tested, 26 failed conversion due to incomplete code, missing documentation, or incompatible dependencies. However, Zou views these failures as useful signals. "Agentification itself is a useful certificate that says, 'This work is relatively complete and well documented,'" he notes.

Olivier Elemento, who directs the Englander Institute for Precision Medicine at Weill Cornell Medicine and peer-reviewed the study, sees broader implications: "It's 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."

Zou envisions a future where papers routinely include "agent availability" statements alongside traditional data and code availability declarations—creating virtual corresponding authors available to answer questions in any language, at any time.

The researchers have already applied the concept recursively, converting their own Paper2Agent manuscript into an agent accessible at paper2agent.ai. Details were first reported by IEEE Spectrum.