Stanford's Paper2Agent turns research papers into AI agents

Stanford researchers have built a system that turns a scientific paper, code and data included, into an AI agent that answers questions about it in plain language. The first working prototype spun one up from the AlphaGenome paper in about 45 minutes for $14 of compute, Nature reported on September 16, 2026.

The system, called Paper2Agent, deposits a paper's main text, code, datasets and supplementary materials onto an MCP server, then lets autonomous AI agents build tools that apply the paper's methods to fresh data. Scientists connect through the large language model of their choice. On the AlphaGenome test, the resulting agent answered genetics questions with near-perfect accuracy and outscored Biomni, an academic biomedical AI tool that draws from dozens of databases.

"Knowledge should not be static records," says James Zou, the Stanford computer scientist who led the work. "It really should be dynamic and interactive," with benefits including reproducibility and "all sorts of new kinds of discovery."

Scaled up, Zou's group ran Paper2Agent against 100 computational biology articles from bioRxiv and converted 74 into working agents. Of 599 tools the system generated, 593 passed automated validation, and the agents answered a 300-question benchmark correctly 91.2% of the time, with a margin of 1.6 points. In one case, the AlphaGenome agent independently proposed alternative causal genes for cholesterol-related genetic variants, re-examining a published conclusion without new experiments.

The paper also says agents built from different papers can collaborate autonomously across scientific disciplines. It landed in three of our tracked researcher feeds.