Paper2Agent turns research papers into callable AI agents

Nature has published a paper arguing that research papers themselves should become AI agents. Not summaries of papers, not agents that read papers. The paper as an executable, callable object.

The system is called Paper2Agent, from Jiacheng Miao, James Zou, Jonathan K. Pritchard and colleagues. It runs a six-step pipeline that locates a paper's codebase, sets up a virtual environment, discovers its tutorials, runs them end-to-end, converts them into parameterised tools exposed via Model Context Protocol servers, then validates each tool against the paper's own reported outputs. "Each tool used by a paper agent is validated against the reference codebase's reported results and figures using example datasets, and then locked to ensure reproducibility," the authors write.

The numbers they report are the striking part. Applied to AlphaGenome, Paper2Agent generated 22 MCP tools in about 45 minutes for $14, and hit 98.7 ± 1.3% accuracy on tutorial-derived queries and 100% on novel ones. Scaled to 100 computational biology papers, it agentified 74 of them; of 599 proposed tools, 593 passed validation. On 300 tutorial-derived questions the agents scored 91.2 ± 1.6%, against 86.3% for a Claude + Repo baseline on Sonnet 4.6, at roughly half the per-query cost ($0.20 vs $0.38) and a third of the latency (1.6 minutes vs 4.3 minutes).

The framing the authors want to plant is administrative. "Just as many journals now require data and code availability sections, we anticipate emergence of agent availability sections," they write, a bid to make agentification a norm of publication rather than a novelty. Five researchers on our Who's Who tracker have shared the paper. The authors also describe agents talking to agents: in a psoriasis-genetics case, the AlphaGenome agent flagged GPR137 as the top affected gene, with a follow-up screen showing Spearman correlation 0.613 (p = 3.79 × 10⁻³) under stimulation.

The pipeline is built on the Claude Code SDK and MCP, with a validation loop capped at six attempts per function and a 3% numerical tolerance for passing. The 74% success rate on computational biology is the other side of the same fact: 26 of the 100 papers were not successfully converted, and the abstract offers no breakdown of why.