Paper2Agent turns Nature papers into interactive AI agents

A team led by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou has built a system that packages a published research paper as an AI agent you can call from a chatbot. The framework, Paper2Agent, appears in Nature on 16 September 2026.

The system wraps a paper's manuscript, code, data and workflows behind a Model Context Protocol server that a chat agent such as Claude Code can call. "Each agent serves as an interactive expert on the corresponding paper, capable of demonstrating, applying and adapting its methods to new projects," the authors write.

The numbers. On the AlphaGenome benchmark, the resulting agent reached 98.7% accuracy on tutorial-derived queries against 82.7% for Claude working directly against the same repository. On novel queries it hit 100%, at roughly 1.9× the median runtime of the baseline.

At scale, the pipeline was pointed at 100 computational biology papers and successfully agentified 74 of them, producing 599 proposed tools of which 593 passed automated validation. Across 300 benchmark questions the ensemble scored 91.2%. In one demonstration the authors chained three agents (AlphaGenome, MPRA-scCRISPRi and Perturb-seq) to identify GPR137 as a causal gene for the psoriasis-associated variant rs887314.

The pipeline runs in six steps: codebase identification, environment setup, tutorial discovery, tutorial execution, tool extraction and MCP server assembly. Tools are "validated against the reference codebase's reported results" and locked before deployment, an attempt to blunt the code hallucinations that dog LLM-generated implementations.

The paper does not report why the papers that failed to convert did so. Five researchers we track have posted the link.