A team at Stanford University has unveiled a framework called Paper2Agent that converts scientific papers and their supporting materials into autonomous AI agents capable of interpreting, reproducing, and extending research findings. Described in a Nature paper published on Wednesday, the system processes full research outputs — including manuscripts, code, datasets, and analysis workflows — to create interactive agents that can answer questions about a study, run experiments, and apply methods to new data.

James Zou, a Stanford professor of biomedical data science and co-author of the study, explained that traditional scientific papers remain static records, but Paper2Agent enables them to function as "virtual authors" with direct access to their methodological tools. The framework establishes a Model Context Protocol (MCP) server that exposes a paper’s technical components, allowing large language models to connect via natural language queries and execute tasks autonomously.

According to the research, the MCP server may operate remotely or locally to safeguard sensitive information, though data transmitted to external AI models must still be handled with caution. Researchers handling protected data are advised to exclude such information from the system. The authors emphasize that Paper2Agent is designed to augment — not replace — human oversight, noting that all outputs should be validated for accuracy to mitigate risks associated with AI hallucinations.

Testing conducted by the team evaluated the framework across 136 papers, including 100 from computational biology. Of these, 74 were successfully transformed into functional agents, with failures attributed to incomplete codebases, missing documentation, or unresolved technical dependencies. The system is open source and available via GitHub, with a live demonstration online that illustrates its capabilities and allows users to test its functionality on additional papers.

Zou indicated plans to develop an online platform where paper-derived agents could collaborate on interdisciplinary research challenges. The team views the technology as a means to enhance reproducibility, accessibility, and iterative progress in scientific inquiry, while maintaining rigorous validation standards throughout the process.