Stanford's Paper2Agent turns papers into callable AI agents

Paper2Agent, a framework from a Stanford-led team, converts a published research paper into an AI agent that behaves as what the authors call a "virtual corresponding author": one that answers questions about the manuscript, runs its methods on new datasets, and can collaborate with agents built from other papers. The system was described on 16 September 2026 in Nature, with authors including Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou.

Applied to more than 100 computational biology papers, Paper2Agent produced working agents from 74 of them. It generated 599 proposed tools; 593 passed automated validation. On its showcase case, an agent built from the AlphaGenome paper produced 22 MCP tools in about 45 minutes at a compute cost of $14, with no human in the loop. Across the benchmark set the agents averaged 91.2 ± 1.6% accuracy on tutorial-style questions, and on AlphaGenome the resulting agent outperformed both standard Claude given access to the codebase and Biomni, a specialist AI co-scientist tool.

"Knowledge should not be static records...it really should be dynamic and interactive," Zou told IEEE Spectrum. He also framed the pipeline itself as a filter on the field: "Agentification itself is a useful certificate that says, 'This work is relatively complete and well documented.'" Twenty-six of the 100 papers failed conversion, mainly because of incomplete code, missing documentation, or incompatible software packages.

In one multi-agent demonstration, three linked agents investigated the genetic basis of psoriasis, converged on GPR137 as a potential causal factor, and proposed ten validation approaches. Dongping Chen, a computer scientist at the University of Maryland, College Park, called the approach "quite compelling," describing the idea as making papers "more dynamic and executable through an agentic interface." The Paper2Agent paper has itself been converted into an agent, hosted at paper2agent.ai.