Stanford's Paper2Agent turns papers into working AI agents

Two AI agents, one built from a genome-mutation prediction study and another from an ADHD genome-wide association study, independently flagged the same molecular variant near the MPHOSPH9 gene as tied to elevated ADHD risk. According to the Stanford group behind the experiment, that link had not been reported in the scientific literature before.

The agents were assembled by a framework called Paper2Agent, described by James Zou's team in Nature on 16 September. It wraps each paper's manuscript, code, data and workflows behind a Model Context Protocol server, so a chat agent such as Claude Code can call the paper as a set of tools. Applied to the AlphaGenome study, Paper2Agent produced 22 validated tools in about 45 minutes at a compute cost of roughly $14 on a personal laptop.

Across 100 computational biology preprints from bioRxiv, 74 were successfully agentified. The pipeline generated 599 candidate tools, of which 593 passed automated validation, and the resulting agents scored 91.2% on a 300-question benchmark at about $0.20 per query. Twenty-six papers failed conversion outright.

'Knowledge should not be static records. It really should be dynamic and interactive,' Zou told IEEE Spectrum. He framed the conversion itself as a signal about the underlying work: 'Agentification itself is a useful certificate that says, "This work is relatively complete and well documented."'

Olivier Elemento, a computational biologist at Weill Cornell Medicine, called it 'a real advance in terms of how we think about the publication process, with AI at the center and in a way that makes publications more interactive.' Three researchers we track shared the piece the day it landed.

The MPHOSPH9 result still awaits independent replication, and the retrieved coverage does not spell out what tripped up the 26 papers Paper2Agent could not convert.