Stanford’s Paper2Agent Reimagines Scientific Papers as Interactive AI Agents

(Chokniti-Studio/Shutterstock)

A new framework proposes a way to turn scientific papers into AI agents that can be queried, tested, and reused for new experiments. The system, called Paper2Agent, is described in a preprint paper from Stanford University titled “Reimagining Research Papers as Interactive and Reliable AI Agents.” The authors make the case that papers containing computational methods should not remain static records. Instead, they should become living tools that anyone can interact with and verify.

For decades now, reproducibility has been a persistent problem in computational science. Many studies now include links to code repositories, but reproducing results often involves dealing with outdated software, missing documentation, and differences in computing environments. Paper2Agent aims to eliminate these barriers by automatically converting a research paper’s code into a self-contained, interactive system accessible through natural language. Rather than reading a paper and trying to rebuild its workflow, a scientist could simply ask the paper’s “agent” to perform an analysis or explain a method.

Building an Agent From a Paper’s Code

Paper2Agent begins by locating the code repository associated with a paper and setting up a clean computational environment that mirrors the original conditions. It then identifies key functions or workflows within the code and turns them into discrete “tools” with defined inputs and outputs. Automated tests are generated to confirm that these tools reproduce the same results as the original publication. Once verified, the tools are served through the Model Context Protocol, allowing language models to call the methods directly.

When connected to an AI front end, such as Claude Code, the result is an interactive system. The authors give the example of a user asking the system to “Apply the method in this paper to the newly generated dataset,” and the agent will automatically run the pipeline, produce results, and present interpretable outputs, they say. The authors propose that through abstracting away technical details, the agent lowers barriers to method adoption, ensures reproducibility, and helps researchers focus on insights rather than implementation. The framework is designed to handle both simple queries and complete analytical workflows, always within the verified environment that matches the original study.

Overview of the Paper2Agent. (A) Paper2Agent turns research papers into interactive AI agents by building remote MCP servers with tools, resources, and prompts. Connecting an AI agent to the server creates a paper-specific agent for diverse tasks. (B) Workflow of Paper2Agent. It starts with codebase extraction and automated environment setup for reproducibility. Core analytical features are wrapped as MCP tools, then validated through iterative testing. The resulting MCP server is deployed remotely and integrated with an AI agent, enabling natural-language interaction with the paper’s methods and analyses. (Figure and Caption Credit: Paper2Agent Authors)

Testing the Framework on Real Research

To test the idea, the Stanford authors applied Paper2Agent to three bioinformatics papers: AlphaGenome, a model for interpreting genomic variants; TISSUE, a method for spatial transcriptomics; and Scanpy, a popular toolkit for single-cell RNA sequencing analysis.

The system successfully reproduced published results from each paper and could handle novel queries that went beyond the original tutorials. In one case, the AlphaGenome agent reinterpreted a genetic variant and suggested a different causal gene than the authors had proposed. This finding highlights the potential for Paper2Agent to not only reproduce results but also to reexamine scientific claims with the same data and methods.

According to the paper, the full conversion process can be completed in a matter of hours using standard computing resources. Each resulting agent runs inside a controlled environment with its own test suite, which helps prevent the types of model hallucinations that can occur when LLMs attempt to generate code from scratch.

Challenges Remain on the Road to Reproducible Science

The promise of Paper2Agent is in how it merges reproducibility with accessibility. By embedding scientific methods in an interactive agent, researchers who are not expert programmers can still test ideas, validate results, and build upon published work. For fields such as computational biology, where software complexity can hinder replication, this could lower barriers and encourage more transparent science.

(mechichi/Shutterstock)

There are some caveats and challenges left to contend with, the authors recognize. Many research codebases are messy or incomplete, and automated extraction may fail without human oversight. Maintaining compatibility as dependencies evolve will also require active curation. The authors acknowledge that Paper2Agent is still a proof of concept, and that scaling it across disciplines will depend on how consistently scientists share their data and code.

Paper2Agent hints at a future where reading a scientific paper could mean interacting with it in new ways. In place of code repositories and readme files, future studies might include interactive agents capable of explaining, reproducing, and extending the work they describe. Whether that vision takes hold will depend on how researchers adopt and maintain such systems. If successful, the idea could help close one of the biggest gaps in computational research: the distance between what is published and what can actually be reproduced. Read the full preprint at this link.

Related