Stanford's Paper2Agent turns research papers into AI agents

Stanford University researchers have introduced Paper2Agent, an open-source framework designed to transform static academic papers and their accompanying codebases into interactive, runnable AI agents. Described in a Nature study published on September 16, the system goes beyond simply answering questions about a paper. Instead, it automatically extracts core workflows to build a tested, runnable toolkit that can execute the described methods on new datasets.

To demonstrate the system, the team led by computer scientist James Zou focused on computational biology. They fed Paper2Agent the documentation and code for AlphaGenome, a deep-learning model that predicts how DNA mutations affect gene regulation. Operating on a personal laptop, the framework spent 45 minutes and less than $15 in computing costs to generate 22 distinct tools. A built-in testing agent validated each tool against reference results, attempting to diagnose and fix errors up to six times per function. The validated tools were then packaged into a Model Context Protocol server and connected to Claude Code, creating an agent that outperformed both standard Claude and the specialist AI tool Biomni.

The researchers also demonstrated how these agents can collaborate. By linking the AlphaGenome agent with two other paper-derived agents, the system investigated the genetic basis of psoriasis. The combined agents identified a gene called GPR137 as a likely cause and proposed 10 validation methods. When tested, silencing GPR137 produced gene activity changes matching those of psoriasis-linked variants.

While the tool offers immense potential for making research reproducible, it also highlights existing gaps in scientific publishing. Out of 100 computational biology papers the researchers tested, 26 failed to convert into agents because of missing documentation or broken code. For practitioners, Paper2Agent provides a way to bypass the tedious process of debugging undocumented repositories, while offering a new standard for verifying that published research is complete and reproducible.