Stanford's Paper2Agent turns any paper into an AI agent

Stanford's Paper2Agent built a working AI agent from the AlphaGenome paper in about 45 minutes at a computing cost of US$14, Nature reports.

The system, published in Nature on 16 September 2026, feeds a paper's main text, code, datasets and supplementary materials onto a Model Context Protocol (MCP) server. A team of AI agents then autonomously builds tools that reapply the paper's methods to fresh data, and scientists interact with the resulting agent through whichever large language model they prefer.

On its AlphaGenome test the agent answered genetics questions with 'near-perfect accuracy' and outscored Biomni, a biomedical AI drawing from dozens of databases. The Paper2Agent version also flagged an alternative gene explaining cholesterol-related genetic variants, which the authors present as an example of an agent that can 're-evaluate published conclusions without designing entirely new experiments.'

The Stanford team, led by James Zou, frames the output as a 'virtual corresponding author.' In a post on X, Zou wrote that 'Scientific knowledge is traditionally stored in passive papers' and that Paper2Agent turns them into authors that 'answer questions, apply its methods and collaborate w/ other paper agents to make new discoveries.' He is careful about credit: 'It's still important to attribute the final discoveries and reference them back to original papers and original human authors.'

Nature's write-up does not publish head-to-head accuracy figures against Biomni, and the reporting details a single demonstration, AlphaGenome, rather than a broader benchmark set.