Paper2Agent turns research papers into interactive AI agents

Paper2Agent, published in Nature by Jiacheng Miao, James Zou and colleagues, converted 74 of 100 computational biology papers into working AI agents that answer natural-language questions about the underlying methods. Across 300 benchmark questions the agents scored 91.2% ± 1.6%, against 80.3% for Claude Code operating directly on the same repositories.

The framework builds an MCP (Model Context Protocol) server for each paper. Sub-agents scan tutorials, extract callable tools, spin up an execution environment, and validate outputs against the paper's own reference numbers before deployment, typically on Hugging Face Spaces.

In the authors' framing, papers become "virtual corresponding authors" that expose manuscripts, code, datasets and workflows as executable tools, turning a static PDF into "active systems that accelerate use and discovery."

The specific numbers do the work. An AlphaGenome-based agent produced 22 validated tools in roughly 45 minutes at $14 in compute, hit 98.7% ± 1.3% accuracy on tutorial-derived queries and 100% on novel queries, and ran 1.9 to 3.1 times faster than working through the raw repository. A Scanpy agent for single-cell analysis was built in about 45 minutes for $13.

The floor is uneven. About 26% of the computational biology papers the team tried to convert failed the pipeline, blamed on missing code, missing data, or environments that would not build. Tests verify numerical accuracy against a 3% tolerance and figure matching against perceptual hashing.

The authors also propose that journals adopt an "agent availability" section alongside code and data availability, and describe multiple paper agents collaborating to nominate GPR137 as a probable causal gene for psoriasis at the rs887314 locus. Five researchers we follow shared the link the same day it went up.