Paper2Agent turns research papers into runnable AI agents

Paper2Agent, described in Nature by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou, converts published research papers into interactive AI agents that expose the paper's code and methods behind natural-language queries. The framework was tested across three corpora totaling 136 papers spanning computational biology, data science, AI, statistics, and astrophysics.

On a run of 100 computational biology papers, the system "successfully agentified 74 of 100 computational biology papers, generating 599 proposed tools with 593 passing validation." The authors report a median runtime 1.9× faster than direct repository access, at approximately $14 per paper to create a functional agent.

The showcase is an AlphaGenome agent that hit 98.7 ± 1.3% accuracy on tutorial-derived queries and 100.0% on novel queries. A user can ask it to "Interpret the expected effect of this variant on chromatin accessibility in muscle cells" without touching the setup.

Under the hood the framework leans on the Model Context Protocol, which the paper says "allows structured APIs and tools to be exposed in a way that is directly accessible to LLMs." Six specialized sub-agents divide the work: an environment manager, a tutorial scanner, a tutorial executor, a tool extractor-implementor, a test verifier-improver, and an orchestrator.

One worked example goes further. Three paper-derived agents collaborated across sources to identify GPR137 as the probable causal gene for a psoriasis-associated genetic variant.

The authors propose adding an "agent availability" section to research papers alongside the existing code and data requirements, treating agents as maintained research artifacts. Five researchers in our Who's Who tracker had shared the paper's link by the time we wrote this up.