Paper2Agent turns research papers into working AI agents

Paper2Agent, an automated framework introduced in Nature on September 16, converted 74 of 100 computational biology papers into functional AI agents. The framework, from Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard, and James Zou, treats a published paper as something to be run.

From the abstract: Paper2Agent converts a paper "into an AI agent that functions as a virtual corresponding author, exposing its manuscript, supplementary materials, datasets, code and workflows as active, agent-native knowledge rather than static text." It builds a Model Context Protocol server for each paper by running a multi-agent pipeline over the codebase and its tutorials, then generates and runs tests to firm up robustness.

Three case studies do the persuading. The team built agents from AlphaGenome (genomic variant interpretation), Scanpy (single-cell analysis), and TISSUE (spatial transcriptomics). On the AlphaGenome build, Paper2Agent produced 22 tools in about 45 minutes for roughly $14, and hit 98.7 ± 1.3% accuracy on 15 tutorial-derived queries. In a joint task, multiple agents were chained to "prioritize a causal gene for psoriasis", nominating GPR137.

The complaint the paper starts from is old. "Papers are fundamentally passive objects," the introduction argues; "a reader must discover the paper... parse its contributions and manually determine how to apply them to their own work." Paper2Agent's answer is to hand the reader something callable in natural language.

The work already found its readership: five of the researchers we track posted the Nature link when it landed. Twenty-six of the hundred computational biology papers did not agentify successfully, and the abstract does not itemize why, or who maintains an MCP server as the underlying chat models drift.