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

Stanford researchers report in Nature that their Paper2Agent framework converted 74 of 100 computational biology papers into functional AI agents, generating 599 tools of which 593 passed validation. On a 300-question benchmark the resulting agents scored 91.2% accuracy, against 80.3% for Claude Code operating on the same repositories directly. Queries ran at roughly $0.20 and 1.6 minutes each.

The framework, built by Jiacheng Miao, James Zou and colleagues, uses Model Context Protocol servers to expose a paper's manuscript, code and workflows as callable tools rather than static text. "By turning static papers into interactive AI agents, Paper2Agent introduces a paradigm for knowledge dissemination and a collaborative ecosystem of AI co-scientists," the authors write.

The most concrete case study chains three paper agents together. An AlphaGenome agent, an MPRA-scCRISPRi agent and a Perturb-seq agent — each spun up from its own paper — jointly identified GPR137 as a causal gene for the psoriasis-associated variant rs887314, with a significant correlation under stimulated conditions (Spearman ρ = 0.613, P = 3.79 × 10⁻³). The AlphaGenome agent alone was generated in about 45 minutes for $14 and hit 98.7% accuracy on tutorial queries and 100% on novel ones.

Coverage is not universal. Twenty-six of the initial 100 computational biology papers did not agentify, and the paper does not attribute the misses to a single cause. The evaluation later expanded to 136 heterogeneous papers spanning biology, AI, statistics and astrophysics, where non-biology papers reached 98.1% accuracy on execution tasks and data or discovery papers 89% on synthesis. Paper2Agent runs on Claude Sonnet 4 and Claude Code, and the AlphaGenome agent is currently hosted on Hugging Face Spaces.

Five of the researchers we track shared the paper into our Who's Who feed, which is a heavier circulation than most methods papers pick up.