Stanford's Paper2Agent turns papers into working AI agents

In about 45 minutes on a personal laptop, for about $14, a system called Paper2Agent turns a research paper into an AI agent that can answer questions about it, run its methods on new data, and collaborate with agents built from other papers. Nature reported the work on 16 September 2026.

The Stanford-led team of Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou describes the output as a 'virtual corresponding author.' The framework parses a paper's text, supplementary materials, code and datasets, deposits them on an MCP server, and lets a team of AI agents build callable tools from the methods.

On the AlphaGenome paper, Paper2Agent produced 22 tools that all passed automated validation and hit 98.7% accuracy on tutorial-derived queries. Applied to 100 computational biology papers from bioRxiv, it turned 74 into working agents comprising 593 tools, scoring 91.2% on a 300-question benchmark at $0.20 per query. Twenty-six papers did not convert.

'Knowledge should not be static records. It really should be dynamic and interactive, and this has many benefits, including making knowledge more reproducible but also enabling all sorts of new kinds of discovery,' Zou told IEEE Spectrum. He also frames the conversion itself as a quality signal: 'Agentification itself is a useful certificate that says, "This work is relatively complete and well documented."'

Documented limitations: incomplete code in source materials, missing documentation, and incompatible software package dependencies. Dongping Chen at the University of Maryland called the idea of making papers 'more dynamic and executable through an agentic interface' quite compelling; Olivier Elemento at Weill Cornell Medicine called it 'a real advance in terms of how we think about the publication process, with AI at the center.' Three researchers in our Who's Who directory posted the paper's link the same week it appeared.