Paper2Agent turns static research papers into active AI agents

A team led by Jiacheng Miao and James Zou has built a system that ingests a research paper together with its codebase and produces a working AI agent that answers questions about the paper's methods in natural language. Their Nature paper describes Paper2Agent, a framework that agentified 74 of 100 computational biology papers, with 593 of 599 auto-generated tools passing validation.

The mechanism is the Model Context Protocol. Paper2Agent parses a manuscript and its repository, wraps the methods as MCP tools, generates and runs tests, then serves the whole thing to a downstream chat agent. In the case studies that agent is Claude Code, powered by Claude Sonnet 4. On AlphaGenome, a genome-scale foundation modeling paper, the pipeline produced 22 MCP tools in 45 minutes for $14 in compute. The resulting agent hit 98.7 ± 1.3% accuracy on tutorial queries and 100.0 ± 0.0% on novel queries, and rejected 100% of permuted paper-question pairs that were out of scope.

On a broader benchmark of 300 tutorial questions, Paper2Agent scored 91.2 ± 1.6% versus 80.3 ± 2.3% for a Claude + Repo baseline, at $0.20 per query against $0.38 for direct repository access. It ran 1.9× and 3.1× faster than Claude + Repo and Biomni respectively.

Five of the researchers we follow posted the link the same week it appeared, which tracks with the authors' framing. "Paper2Agent reimagines research dissemination by turning static papers into active AI agents," they write, going further to argue: "The ease with which a paper can be transformed into an agent may itself serve as a practical measure of reproducibility."

Case studies also cover Scanpy and TISSUE, and one demo has multiple paper agents collaborating to prioritize a psoriasis causal gene by integrating computational predictions with experimental validation data.