Stanford's Paper2Agent turns papers into interactive AI agents

Paper2Agent, a system from Stanford published in Nature on Sept. 16, converts a research paper, including its text, code and data, into an interactive AI agent that answers questions about the work and reruns its methods on fresh datasets. Building an agent for the AlphaGenome paper took roughly 45 minutes and about $14 in compute.

The tool deposits a paper's contents onto a Model Context Protocol server, which any MCP-compatible chat model like Claude Code can call. In the team's tests, the AlphaGenome agent answered genetics questions with 'near-perfect accuracy' and outperformed Biomni, an academic biomedical AI tool with access to dozens of databases. Rerun on cholesterol-related DNA changes, the agent nominated a different causal gene than the original AlphaGenome paper had.

"For essentially all of human history, the way that we represent knowledge is in the form of these very passive artifacts," James Zou, the Stanford computer scientist who co-authored the work, told TechXplore. He said the technology "can help us to reimagine what knowledge looks like in the future."

The more provocative demonstration comes from pairing two paper agents. When the AlphaGenome agent was connected to one built from an ADHD genome-wide association study, the pair flagged a molecular variant near a gene called MPHOSPH9 as linked to increased ADHD risk, a connection Zou said had not been reported before. No independent replication is described.