Paper2Agent turns research papers into working AI agents

Paper2Agent, a framework unveiled in Nature on 16 September 2026, converts computational research papers into interactive AI agents that expose a manuscript's code, data, and workflows through natural-language queries. The system builds Model Context Protocol (MCP) servers from a paper's codebase using multiple analysis agents, auto-generates and validates tests, then exposes the result to chat clients such as Claude.

In evaluations, the AlphaGenome agent scored 98.7 ± 1.3% accuracy on tutorial-derived queries and 100% on novel queries, versus 82.7% for a direct-repo Claude baseline, and ran roughly 1.9× faster. Across 100 computational biology papers, 74 were successfully agentified, 593 of 599 proposed tools passed validation, and a 300-question benchmark landed at 91.2 ± 1.6% accuracy.

Cost figures are modest. Building the AlphaGenome MCP server ran about 45 minutes and roughly US $14; Scanpy came in at US $13. Each query afterwards costs about US $0.20 versus US $0.38 for direct repository access.

The authors write that "Paper2Agent reimagines research dissemination by turning static papers into active AI agents. Each agent serves as an interactive expert on the corresponding paper, capable of demonstrating, applying and adapting its methods to new projects." Automated tests verify that "expected files are generated, numerical results fall within tolerance thresholds and figures match references." A multi-agent case study points to GPR137 as a candidate causal gene for psoriasis, arrived at by combining AlphaGenome predictions with CRISPR and Perturb-seq data.

Five researchers on our Who's Who tracker posted the source link within days of publication.