Paper2Agent turns research papers into interactive AI agents

A framework called Paper2Agent, described in Nature on 16 September, converts research papers into interactive AI agents by wrapping their code and methods in Model Context Protocol servers. Applied to 100 computational biology papers, the system successfully converted 74; of 599 auto-generated tools across those runs, 593 passed automated validation.

The demo agents carry the specifics. An AlphaGenome agent produced 22 MCP tools in about 45 minutes for roughly $14 on a personal laptop, scoring 98.7 ± 1.3% on tutorial questions and 100.0 ± 0.0% on novel queries, against a 'Claude + Repo' baseline of 82.7% and 78.7%. A Scanpy single-cell agent was built with 7 tools at similar time and cost. Across 300 tutorial-derived questions on the wider sample, Paper2Agent averaged 91.2 ± 1.6% accuracy versus 80.3 ± 2.3% for the baseline (p < 0.0001), and the authors report it ran 34× cheaper and 15× faster than a browser-based baseline on discovery-style papers.

The authors frame this as a rewrite of what a paper is. 'Paper2Agent reimagines research dissemination by turning static papers into active AI agents. Each agent serves as an interactive expert,' they write, and propose that journals add an 'agent availability' section alongside today's data and code statements. Five researchers in our Who's Who tracker posted the link on publication day.

Not every paper cooperated. Twenty-six percent of the biology sample failed to agentify because of missing code, incomplete documentation, or environment configurations the pipeline could not resolve. In a joint test, three linked agents (AlphaGenome plus two experimental-screen agents) converged on GPR137 as a probable causal gene for the psoriasis-associated variant rs887314. The authors describe this as 'a new mode of scientific communication, moving beyond static dissemination to interactive collaboration.'