Paper2Agent turns research papers into interactive AI agents

Jiacheng Miao, James Zou and colleagues want your next research paper to answer questions in natural language. In a paper published September 16 in Nature, the team introduces Paper2Agent, a framework that reads a manuscript along with its supplementary materials, datasets, code and workflows and exposes them through Model Context Protocol servers as tools an AI agent can call.

The pipeline is itself multi-agent. Specialized sub-agents handle environment setup, tutorial discovery, tool extraction, and testing, and any tool that passes validation is locked to preserve reproducibility. On the systems the authors built by hand, the AlphaGenome agent hit 98.7% accuracy on tutorial-derived queries and 100% on novel ones, and the Scanpy agent reproduced human researcher results across diverse datasets. Scaled across 100 computational biology papers, 74 were successfully turned into working agents, with 91.2% accuracy on benchmark questions.

The authors go further than lookup. Chaining three paper agents in what they call "cross-screen, cross-modality signature-correlation" analysis, the group identified GPR137 as a probable causal gene for psoriasis.

"Paper2Agent transforms research output from passive artefacts into active systems that accelerate use and discovery," the authors write, arguing that a paper's manuscript, supplementary materials, datasets, code and workflows should live as "active, agent-native knowledge rather than static text." They frame the system as a tool for augmenting scientific discovery, with human researchers still in the loop.