Paper2Agent turns 74 of 100 bio papers into working AI agents

Stanford researchers converted 74 of 100 computational biology papers into working AI agents at a median cost of $14 apiece, running in about 45 minutes on a personal laptop. The tool, Paper2Agent, was described in Nature on 16 September 2026 by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou, who frame each output as a 'virtual corresponding author': a chat interface that answers questions about the paper, reruns the paper's methods on new data, and can hand off to agents built from other papers.

'Knowledge should not be static records,' Zou told IEEE Spectrum. 'It really should be dynamic and interactive.'

The mechanics: Paper2Agent reads the manuscript and its accompanying codebase, packages both onto a Model Context Protocol (MCP) server, then generates and runs its own tests to harden the resulting tools. In the 100-paper trial, 593 of the 599 tools it produced passed automated validation. On 300 benchmark questions drawn from the papers' own tutorials, the agents answered with 91.2% ± 1.6% accuracy, against 80.3% for a Claude-plus-repository baseline. Applied to the AlphaGenome paper, it built 22 tools in roughly 45 minutes at about $14 in compute.

The stumbles are equally concrete. 26 of the 100 papers failed to convert. And the benchmark that produced the headline accuracy number was built from the papers' own tutorials, a friendly test.

Outside biology, Zou called agentification 'a useful certificate that says, this work is relatively complete.' Dongping Chen of the University of Maryland described the approach as 'quite compelling' for making papers 'more dynamic and executable through an agentic interface,' and Weill Cornell Medicine's Olivier Elemento called it 'a real advance in terms of how we think about the publication process, with AI at the center.' Three of the researchers on our tracker circulated the link the week the paper dropped.