Stanford researchers released a new framework Wednesday that converts scientific papers into interactive AI agents capable of reproducing analyses, answering questions about findings, and even collaborating with other paper agents on new research problems. The tool, called Paper2Agent, addresses a persistent bottleneck in science: the gap between a published paper's static text and the working code, data, and workflows that produced its results.
"Papers have been static documents for centuries," said James Zou, a Stanford computer scientist and biomedical data science professor and one of the paper's authors. "Paper2Agent turns them into active AI agents that can answer questions, apply their methods, and collaborate with other papers to make new discoveries."
Giving a large language model access to a scientific paper produces unpredictable results, Zou added. His team wanted an agent that could act as a "virtual author" with hands-on experience of a paper's work, not just a model that reads and attempts to understand it. The framework, described in a paper published in Nature, uses a paper's manuscript, supplementary materials, code, datasets, and executable examples to build a MCP (Model Context Protocol) server that exposes the research's tools and workflows. An LLM agent then connects to this server and accepts natural-language requests to run demonstrations, apply methods to new data, or reproduce results.
How the agentification works
The MCP server can be hosted remotely or run locally to protect sensitive information, though any data sent through the system still reaches whichever LLM backend Paper2Agent connects to. Zou told The Register that users with protected health information or other sensitive data should exclude it from the tool. Paper2Agent should work with any AI coding agent, he said, though not all have been tested.
Each tool used by a paper agent is validated against the original paper's results and figures, then locked to reduce hallucination risk and minimize randomness. The researchers still emphasize that users should evaluate everything the agent produces. The paper frames Paper2Agent as a tool for augmenting scientific discovery and improving access, reproducibility, and reuse, not as an autonomous or authoritative source of scientific conclusions.
Early testing shows promise and limits
The team evaluated Paper2Agent across 136 papers in three groups, including 100 computational-biology papers. Of those 100, 74 were successfully converted into agents. Failures largely stemmed from incomplete codebases, missing documentation, or environment configurations that could not be resolved. This suggests the tool's effectiveness depends heavily on the quality and completeness of a paper's associated research outputs.
Paper2Agent is open source and available on GitHub. A live version online can explain the Paper2Agent paper itself and reproduce its results, and Zou said it can also ingest other papers to demonstrate how it works on different projects. The team's next goal is an online platform where paper agents can collaborate and discuss agentified scientific discoveries.
Why this matters for science and research professionals
Paper2Agent shifts the unit of scientific communication from a static PDF to an executable, queryable system. For researchers who regularly need to verify claims, adapt methods to new datasets, or understand how a paper's analysis actually runs, this cuts through the friction of digging through supplementary materials and undocumented code. The 74 percent success rate on computational-biology papers indicates the approach works best when code and data are already well-organized - a standard worth meeting for labs that want their work to be reproducible and reusable. As tools like this mature, the expectation that published research should include working, agent-ready workflows may reshape norms in AI for science and research.
Your membership also unlocks:
