Stanford's Paper2Agent turns research papers into AI agents

Stanford Medicine researchers have built a system that turns a research paper and its codebase into a functioning AI agent a reader can query in plain English. Writing in Nature, Jiacheng Miao, James Zou and colleagues describe Paper2Agent, an automated pipeline that constructs Model Context Protocol servers from a paper's manuscript, data, and code, then generates and runs its own tests to harden the resulting agent.

"Paper2Agent transforms research output from passive artifacts into active systems that can accelerate downstream use, adoption, and discovery," the paper's abstract states, framing conventional publications as barriers to reuse when readers have to unpick a codebase to apply the methods themselves.

The prototype agentified 74 of 100 bioRxiv computational biology papers. Of 599 proposed tools, 593 passed validation. Across 300 questions drawn from those 100 papers, the resulting agents scored 91.2 ± 1.6% accuracy, ahead of a baseline that gave Claude direct repository access and scored 80.3%. A hand-built AlphaGenome agent, produced in about 45 minutes at roughly $14, hit 98.7 ± 1.3% on tutorial queries and 100% on novel ones, with per-query costs falling to $0.20 against $0.38 for the repository baseline.

The paper also demonstrates a discovery loop. Multi-agent collaboration flagged GPR137 as the probable causal gene at the psoriasis-associated locus rs887314. The team's arXiv preprint of the same work highlighted a parallel use case, nominating a splicing variant tied to ADHD risk.

Code is MIT-licensed and open source on GitHub, with prebuilt AlphaGenome, Scanpy and TISSUE agents hosted on Hugging Face Spaces. Five researchers on our Who's Who tracker posted the source link.