Paper2Agent turns research papers into interactive AI agents

A Nature paper published on 16 September 2026 by Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou describes an automated pipeline that wraps a manuscript's code, tutorials and datasets inside a Model Context Protocol server. The paper then answers natural-language queries by running its own methods.

The authors call this 'turning static papers into active AI agents' that function as 'virtual corresponding authors.' Building the AlphaGenome agent took 45 minutes and about $14, produced 22 MCP tools, and reached 98.7 ± 1.3% accuracy on tutorial-derived queries and 100.0 ± 0.0% on novel ones. A Claude-with-repository baseline scored 82.7 ± 3.4%. A Biomni comparator scored 37.3 ± 4.0%. The paper agent ran 1.9× faster than Claude + Repo.

Scaled up, Paper2Agent processed 136 papers across three corpora with no human intervention. Seventy-four of 100 computational biology papers were successfully agentified. Of 599 proposed tools, 593 passed validation. On 300 tutorial-derived questions the resulting agents averaged 91.2 ± 1.6% accuracy, versus 80.3 ± 2.3% for Claude Code with repository access.

In one multi-agent demonstration, three paper agents worked together to identify GPR137 as the likely causal gene behind the psoriasis-associated variant rs887314, with a Spearman ρ of 0.630 (p = 4.71 × 10⁻³) between CRE perturbation and gene knockdown signatures. 'Paper agents can interact directly with each other' to 'perform analyses, test hypotheses and generate new insights,' the authors write.

Five researchers on our tracker had already shared the link by the time we picked it up.

The pipeline is not universal. The authors report that 'a substantial fraction of repositories still could not be successfully agentified, typically owing to incomplete codebases, missing documentation or unresolvable environment configurations.' They frame Paper2Agent as 'a tool for augmenting scientific discovery' and stress that 'researchers remain responsible for selecting directions and evaluating evidence,' while proposing that journals add an 'agent availability' section alongside data and code.