Stanford's Paper2Agent turns papers into callable AI agents

Stanford researchers built an AI agent from the AlphaGenome paper in about 45 minutes at a computing cost of roughly $14, and it answered genetics questions with what the authors describe as near-perfect accuracy, outperforming an existing biomedical agent called Biomni. The system that produced it, Paper2Agent, is described in Nature. It converts a paper's text, code, datasets and workflows into an MCP server that MCP-compatible agents, including Claude Code, can query in plain language. The Nature news piece publishes no confusion matrix and no model card for the underlying LLM.

The lead author is postdoctoral scholar Jiacheng Miao; the senior author is James Zou, associate professor of biomedical data science at Stanford. "Instead of having only passive artifacts, why don't we convert each static record into an active embodiment of knowledge?" Zou said. He compared the underlying layout to "almost like a filing system," where different sections of a paper live in separate folders that agents can call.

The team has built more than 100 paper agents so far. In one demonstration, they stood up agents for two unrelated papers, a genome-mutation prediction tool and an ADHD genome-wide association study, and let them work together. The pair independently identified a molecular variant near the MPHOSPH9 gene associated with increased ADHD risk, a link the group says is not previously documented in the scientific literature. In a separate test on genes linked to cholesterol, the AlphaGenome agent proposed a different causal gene than the original paper had settled on, which Zou described by saying "AlphaGenome's data on genetic variants support both hypotheses."

Zou wants attribution to survive the transformation. "It's still important to attribute the final discoveries and reference them back to original papers and original human authors," he said, and he acknowledged that the parameters governing agent-to-agent collaboration will need close monitoring. He also floated a longer-term vision of "millions of paper agents" spotting connections across the literature at scale.