Stanford's Paper2Agent turns any paper into a callable AI agent

Stanford researchers have built a system that converts any scientific paper into an interactive AI agent, successfully for 74 of the 100 bioRxiv computational biology papers they tried. In one demonstration, two of those agents combined findings to flag a previously unreported variant near the MPHOSPH9 gene associated with increased ADHD risk.

The tool, Paper2Agent, is described in Nature alongside a paper published September 16, 2026 by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou. The pipeline reads a manuscript together with its data and code, then exposes them as Model Context Protocol servers so any compatible AI agent can call the paper's methods as tools. Nature calls the resulting agents 'virtual corresponding authors' that can answer questions about a paper, apply its methods to new data, and collaborate with agents built from other papers.

On the 100-paper benchmark, the pipeline produced 593 of 599 tools that passed automated validation and scored 91.2% on 300 tutorial-derived questions at roughly $0.20 per query, ahead of an 80.3% baseline where the same model was given direct access to the paper's repository. A dedicated agent built for AlphaGenome, packaged as 22 tools, hit 98.7% on tutorial questions and 100% on novel genomic queries.

'Instead of having only passive artifacts, why don't we convert each static record into an active embodiment of knowledge?' the authors write. Three researchers we track posted the paper the day it appeared.