Presents Paper2Agent, a tool that converts static research papers into dynamic AI agents capable of answering questions and applying methods to new data.
Paper2Agent, developed by researchers at Stanford University including James Zou, is an automated framework that transforms static research papers into interactive, executable AI agents. Published in Nature on September 16, 2026, the system converts a paper’s manuscript, code, data, and workflows into a Model Context Protocol (MCP) server, enabling users to interact with the research via natural language through large language models like Claude Code.
The tool operates by deploying specialized AI agents to autonomously analyze the source code and build validated tools that reproduce the original findings. Key capabilities include: Dynamic Interaction: Acting as a "virtual corresponding author" to answer complex scientific queries and explain methods without requiring programming expertise. Reproducibility: Validating generated tools against original results to ensure accuracy, with benchmarks showing 98.7% accuracy on tutorial queries and 100% on novel queries for the AlphaGenome case study. Cross-Paper Collaboration: Enabling different paper agents to autonomously collaborate; for example, agents for AlphaGenome and ADHD GWAS data jointly identified a novel splicing variant associated with ADHD risk. Efficiency: Building an agent for the AlphaGenome paper took approximately 45 minutes and cost US$14, significantly outperforming existing biomedical AI tools like Biomni.
By shifting scientific knowledge from passive PDFs to active, queryable services, Paper2Agent aims to lower barriers to method adoption and accelerate scientific discovery through a new ecosystem of AI co-scientists.
Paper2Agent: from PDF to executable expertise. The Nature article describes Paper2Agent, a tool that reframes a research paper as a specification for a specialized AI agent rather than merely a static document. In this model, a paper’s methods, assumptions, inputs, outputs, and evaluation criteria are extracted and packaged into an interactive agent that can answer method-specific questions, clarify scope, and apply the described workflow to new data. The emphasis is on turning published scientific knowledge into a more operational form: an agent that can be queried, reused, and potentially coordinated with other agents or tools.
Why it matters. This matters because much scientific knowledge remains locked in prose, figures, supplementary material, and fragmented code that are difficult to reuse outside the original lab or domain. By converting papers into agents, Paper2Agent could lower the barrier to applying published methods, accelerate method transfer, and support more composable AI research workflows. It also points to a broader shift from retrieval-augmented question answering toward action-oriented systems that can not only explain a method but also instantiate or adapt it.
Key caveat: validation and provenance. For a technically literate audience, the central challenge is reliability. The value of such a system depends on how well it handles ambiguous methods, missing implementation details, edge cases, and the risk of hallucinating plausible but incorrect steps. It also raises important questions about citation, auditability, and whether the agent’s outputs can be traced back to the original paper. If these issues are addressed, the approach could improve reproducibility and make research methods more accessible; if not, it risks automating confident but flawed workflows.