Stanford's Paper2Agent turns papers into interactive AI agents

Twenty-two callable tools, forty-five minutes, fifteen dollars in compute. That is what a Stanford team says it took to turn the AlphaGenome paper into a working AI agent that other software can query the way a coder queries an API.

The system, Paper2Agent, was published in Nature on 16 September 2026 by Jiacheng Miao, James Zou and colleagues. The abstract frames it as converting papers 'from passive artefacts into active systems that accelerate use and discovery,' exposing 'manuscript, supplementary materials, datasets, code and workflows as active, agent-native knowledge rather than static text.' The pipeline reads a paper, extracts anything runnable, wraps it as Model Context Protocol servers, and verifies each tool against reference outputs before locking it.

Across 100 computational biology papers the team ran through it, the resulting agents averaged 91.2 ± 1.6% accuracy on tutorial-style benchmark questions, according to IEEE Spectrum. The AlphaGenome agent hit 98.7% on 15 tutorial-derived queries and 100% on 15 novel ones the authors wrote themselves. Twenty-six of the 100 papers failed conversion outright.

'Knowledge should not be static records. It really should be dynamic and interactive,' Zou told IEEE Spectrum. He added a sharper line: 'Agentification itself is a useful certificate that says, "This work is relatively complete."'

Outside voices are interested but hedged. 'It's a real advance in terms of how we think about the publication process, with AI at the center,' said Olivier Elemento of Weill Cornell Medicine. Dongping Chen at the University of Maryland called the idea of executable papers 'quite compelling.'

Nothing in what the team has released explains which 26 papers broke, or why.