Stanford's Paper2Agent turns 74 of 100 papers into AI agents

A Stanford team ran an automated pipeline over 100 bioRxiv computational biology papers, and 74 of them came out the other end as working AI agents. The system, called Paper2Agent, generated 599 tools from those papers; 593 passed automated validation. The resulting agents scored 91.2% across 300 benchmark questions drawn from the papers themselves, against 80.3% for a Claude-plus-repository baseline. The work appears in Nature, published 16 September.

The pipeline converts a paper and its codebase into a Model Context Protocol server. Any MCP-compatible client, such as Claude Code, can then call the paper's methods through natural language. The authors describe each output as a "virtual corresponding author" whose manuscript, supplementary materials, datasets, code and workflows sit on the server "as active, agent-native knowledge rather than static text."

Miao, Zou and colleagues built demonstration agents for AlphaGenome, Scanpy and TISSUE. According to a technical write-up on MarkTechPost, the AlphaGenome agent took about 45 minutes to build for US $14 in compute, and scored 98.7 ± 1.3% on tutorial-derived queries and 82.7 ± 2.4% on open-ended ones.

The paper frames the result as "a new mode of scientific communication, moving beyond static dissemination to interactive collaboration." The tested corpus is 100 papers from a single field, and the 300 benchmark questions were drawn from those same papers, so the accuracy figures leave room for a harder held-out evaluation. Five people in our Who's Who tracker posted the link within days of publication.