Researchers have built an automated system that converts static scientific papers into interactive AI agents capable of running analyses, answering questions, and even collaborating with other paper agents to generate new discoveries. The framework, called Paper2Agent, was described in a paper published in Nature on 16 September 2026 by Jiacheng Miao, Joe R. Davis, and colleagues at Stanford University.
Conventional research papers require readers to manually decipher code, install environments, and adapt methods before they can apply findings to their own work. Paper2Agent addresses this by turning a paper's manuscript, supplementary materials, datasets, and code into what the authors call "agent-native knowledge" - a virtual corresponding author that can execute methods and respond to natural language queries.
How the system works
Paper2Agent uses a multi-agent AI pipeline to automatically transform a paper into a production-ready server. The system first locates and downloads the associated code repository, then extracts reusable functions from tutorial code. A specialized environment agent configures the necessary software dependencies to ensure reproducibility. The result is an MCP (Model Context Protocol) server containing executable tools, static resources like manuscript text and datasets, and prompts that guide the AI through multi-step scientific workflows.
This server connects to a large language model - the team used Claude Sonnet 4 - to create a conversational agent. Users can then interact with the paper's methods through natural language. "In this framework, a paper becomes an executable research artefact that can answer questions, reproduce analyses, apply methods to new data and interoperate with other paper agents," the authors wrote.
Benchmarking against human and AI baselines
The team validated the approach across diverse scientific domains. For the AlphaGenome paper, which describes a computational genomics method, the Paper2Agent-generated agent was benchmarked against human-executed ground truth, Claude Code with direct repository access, and another AI scientist tool called Biomni. Two independent human experts graded the results using predefined rubrics, with 96.7% inter-rater agreement.
For single-cell data analysis, the team created a Scanpy agent that automatically runs preprocessing and clustering workflows. Users only need to provide a data file path. The agent was tested on four publicly available single-cell datasets not included in the Scanpy codebase, and later on seven additional diverse datasets. It "adaptively adjusts parameters based on data characteristics," according to the paper.
To test scalability, the researchers processed three heterogeneous paper corpora without manual cleanup or code modification: 100 computational biology papers, 26 data-focused papers, and 10 non-biology computational papers. For the 26 data- and discovery-focused papers, the resource layer achieved 89.0 ± 3.1% accuracy on 100 synthesis-based questions, outperforming a Claude browser-use baseline at 82.0 ± 3.8%. The major failure modes included missing executable code, missing data or model artifacts, environment failures, and non-generalizable scripts.
Paper agents that collaborate with each other
The system also enables paper agents to interact directly. In one discovery case study, the team connected agents for AlphaGenome, an MPRA-coupled scCRISPRi screen, and Perturb-seq. The AI co-scientist cross-referenced predictions from one paper with experimental data from the other two, autonomously inspecting manuscripts and analyzing supplementary tables to prioritize candidate genes.
"Paper2Agent enables a new mode of AI-driven collaboration in which AI paper agents can interact directly with each other," the authors said. "The agent of a new method paper can autonomously collaborate with the agent of a new data paper to perform analyses, test hypotheses and generate new insights."
Why this matters for science and research professionals
Paper2Agent signals a shift in how research outputs could be structured and shared. The authors anticipate that journals may eventually require an "agent availability" section alongside data and code availability statements, incentivizing researchers to publish their work in formats that can be converted into interactive agents. The framework does not replace human judgment - the authors explicitly frame it as a tool for augmenting discovery and improving access, not as an autonomous source of scientific conclusions.
For computational researchers, this means methods described in papers could become immediately usable without the typical weeks of environment setup and code adaptation. For experimentalists, AI for Science & Research tools like this could lower the barrier to applying complex computational methods to new datasets. The paper is available under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Your membership also unlocks:
