Paper2Agent converts 74 of 100 biology papers to AI agents

Paper2Agent, a framework from a Stanford group led by Jiacheng Miao and James Zou, converted 74 of 100 computational biology papers into working AI agents that answer questions about the methods and rerun them on new data, according to a study published in Nature on 16 September 2026.

The system reads a paper's manuscript, supplementary materials, datasets and code, then packages the validated methods as a Model Context Protocol server that chat agents like Claude Code can call. In the 100-paper trial the pipeline extracted 599 candidate tools and 593 of them passed automated validation. The authors describe the output as exposing a paper as "active, agent-native knowledge rather than static text," and cast each agent as a virtual corresponding author.

On 300 tutorial-derived benchmark questions the agents scored 91.2% accuracy, against 80.3% for direct repository access. An AlphaGenome build produced 22 tools in about 45 minutes on a personal laptop for roughly $14 and hit 98.7% accuracy on 15 tutorial-derived queries. In a separate demonstration three specialist agents collaborated to nominate GPR137 as the causal gene behind a psoriasis-associated variant.

The 26 papers the pipeline could not agentify shared a common pattern: incomplete research artifacts, missing code, unavailable datasets, or scripts too specific to generalize. Five of the AI researchers we track had already circulated the paper on their feeds by the time we picked it up.