Paper2Agent turns research papers into callable AI agents

A team led by James Zou, Jonathan Pritchard and colleagues reports in Nature that they have built Paper2Agent, an automated pipeline that converts a published research paper into an AI agent other agents can call, exposing the paper's code, data, tutorials and workflows through Model Context Protocol servers instead of leaving them buried in a repository.

On tutorial-derived queries, the AlphaGenome agent produced by Paper2Agent hit 98.7 ± 1.3% accuracy and reached 100.0 ± 0.0% on novel queries, against 82.7% and 78.7% for a Claude-plus-repository baseline and 37.3% and 56.0% for the Biomni system, the paper reports. Across 300 tutorial-derived questions spanning several agentified papers, accuracy averaged 91.2 ± 1.6%; on non-biology computational papers, the same pipeline hit 98.1 ± 0.8%.

Scaled to 100 computational biology papers, the framework successfully agentified 74 of them and "generated 599 proposed tools, of which 593 passed automated validation." Building the AlphaGenome agent took roughly 45 minutes and $14 in compute; the Scanpy agent about 45 minutes and $13. Per-query cost ran to $0.20, versus $0.38 for alternatives.

In a proof-of-concept discovery run, the authors chained agents derived from AlphaGenome, an MPRA-coupled scCRISPRi study, and a Perturb-seq of CD4+ T cells to hunt for causal genes in psoriasis. Of the candidates screened, only "GPR137 knockdown showed significant concordance with the CRE perturbation signature," they report.

"Paper2Agent introduces a paradigm for knowledge dissemination and a collaborative ecosystem of AI co-scientists," the authors write. Five researchers we track shared the paper's Nature link within days of its September 16 publication.