Paper2Agent turns static scientific papers into interactive, reliable AI agents

Scientists report in Nature a new system, Paper2Agent, that automatically converts research papers into interactive AI agents—so that instead of reading code, installing dependencies, and manually reconstructing pipelines, researchers can ask a paper to apply its own method to new data. The work, led by James Zou at Stanford University, with first author Jiacheng Miao, introduces a framework that reimagines the paper as an active collaborator rather than a passive document.

Paper2Agent works by analyzing a paper’s manuscript, code, data, and workflows, then using specialized AI agents to build a Model Context Protocol (MCP) server—a standardized interface that lets large language models call the paper’s tools. The framework exposes the paper’s contributions as validated, executable functions and resources, then connects that server to a chat agent such as Claude Code. Users can query the resulting paper agent in plain language, and every tool is checked against reference outputs before deployment, mitigating the risk of AI-generated “code hallucination.”

In benchmark tests, the AlphaGenome agent—created for a genome-scale foundation model—built 22 validated tools in about 45 minutes at a cost of $14. It answered tutorial-derived and novel genomic variant queries with 98.7% and 100% accuracy, outperforming general-purpose coding agents, while reducing runtime. A Scanpy-based agent reproduced human researcher results for single-cell preprocessing and clustering across public datasets, and a third agent was built for the spatial transcriptomics method TISSUE.

The team also showed that paper agents can collaborate. When asked to prioritize a causal gene at a psoriasis-associated locus, agents built from AlphaGenome, an MPRA-coupled scCRISPRi screen, and a Perturb-seq dataset proposed ten candidate strategies, from which researchers selected a cross-screen signature-correlation analysis for execution. The result pointed to GPR137 as the probable causal gene: only GPR137 knockdown significantly matched the regulatory perturbation signature under stimulated CD4+ T-cell conditions (FDR < 0.05).

Not every paper is ready for agentification. In a large-scale test of 100 computational biology papers, 74 could be converted into agents, with 593 of 599 proposed tools passing validation; the main failures were missing code, missing data or model artifacts, and unresolvable environment issues. The authors stress that open-ended scientific reasoning and final interpretation remain human-in-the-loop, and note challenges around security, intellectual property, and maintenance of executable research artifacts.

Still, Paper2Agent demonstrates a practical path from passive publication to reusable scientific tools. The authors suggest that journals may one day include an “agent availability” section, making agent-native artifacts as standard as data and code availability.

Reference: Miao, J., Davis, J.R., Zhang, Y. et al. Reimagining research papers as interactive and reliable AI agents. Nature (2026). https://doi.org/10.1038/s41586-026-11044-y