Paper2Agent Transforms Scientific Papers Into Interactive AI Agents for Collaborative Research

Stanford Researchers Launch Open-Source Framework to Revolutionize Scientific Publishing

Stanford researchers led by postdoctoral scholar Jiacheng Miao and Associate Professor James Zou have introduced Paper2Agent, an open-source AI framework that converts scientific papers into interactive AI agents1

1

2

2

. Published in Nature on September 16, the tool fundamentally reimagines how scientific knowledge is shared and utilized1

1

. Unlike Google's NotebookLM, which simply answers questions about documents, Paper2Agent enables AI agents to actually run the methods described in scientific papers and combine tools from multiple studies1

1

.

The AI framework processes manuscripts along with their associated code, data, and supplementary materials to extract core workflows and create tested, runnable toolkits1

1

. James Zou emphasizes that "knowledge should not be static records" but rather "dynamic and interactive," enabling reproducible research and new kinds of discovery1

1

. The transformation takes approximately 45 minutes on a personal laptop for less than $15 in computing costs1

1

.

How Paper2Agent Converts Scientific Papers Into Functional Tools

Paper2Agent employs a team of AI "worker agents" that analyze published papers and attempt to reproduce the original research from scratch2

2

. The system uses the Model Context Protocol (MCP) to organize and store knowledge, creating a filing system where each paper section lives in separate folders2

2

. The MCP server exposes research tools, resources, and workflows that large language models can access through natural-language requests3

3

.

The framework includes automated validation through a testing agent that runs sub-tools against reference results1

1

. If tests fail, the agent diagnoses problems and attempts fixes up to six times per function before potentially dropping the tool altogether1

1

. In demonstrations using AlphaGenome, a deep-learning model for predicting mutation effects on gene regulation, Paper2Agent produced 22 validated tools covering different aspects of functionality1

1

. These tools were packaged into an MCP server and connected to Claude Code, creating a user-facing agent capable of answering questions in plain English and generating results with visualizations1

1

.

Multi-Agent Research Collaboration Surfaces New Scientific Discoveries

The most compelling capability of Paper2Agent lies in enabling AI agents to collaborate with each other, potentially creating vast research networks2

2

. Stanford researchers demonstrated this by converting two unrelated papers into agents: one on genome prediction tools and another on genome-wide association studies of ADHD risk2

2

. Through agent-to-agent collaboration, the system flagged a molecular variant near the MPHOSPH9 gene associated with increased ADHD risk—a connection not previously reported2

2

.

In another demonstration focused on computational biology, three linked agents investigated the genetic basis of psoriasis1

1

. The collaborative AI agents identified GPR137 as a likely causal factor and proposed 10 validation methods1

1

. When a human researcher selected one approach, the analysis revealed that silencing GPR137 produced gene activity changes strikingly similar to those caused by psoriasis-linked variants in immune cells1

1

. According to Zou, "These agents, because they're able to directly collaborate and communicate, can facilitate all these kinds of collaborations"1

1

.

Reproducible Research Across Diverse Scientific Disciplines

The Stanford researchers tested Paper2Agent across 136 papers in diverse disciplines including statistics, econometrics, astrophysics, and computational biology1

1

3

3

. Of 100 computational-biology papers evaluated, 74 were successfully converted into AI agents3

3

. Failures were attributed primarily to incomplete codebases, missing documentation, or unresolvable environment configurations3

3

.

Dongping Chen, a computer scientist at the University of Maryland, notes that "the idea of making papers more dynamic and executable through an agentic interface is quite compelling"1

1

. The AlphaGenome agent outperformed both standard Claude given the codebase and specialist AI co-scientist tool Biomni when tested with various questions ranging from simple requests to open-ended research problems1

1

.

Addressing AI Hallucinations and Ensuring Research Integrity

While Paper2Agent offers significant potential for reproducible research, concerns about AI hallucinations remain3

3

. The researchers acknowledge that scientists should always evaluate outputs to ensure correctness, viewing Paper2Agent "as a tool for augmenting scientific discovery and improving access, reproducibility and reuse of papers, rather than as an autonomous or authoritative source of scientific conclusions"3

3

.

Each tool undergoes validation against paper results and figures, with tools "locked to ensure reproducibility"3

3

. The MCP server can be hosted remotely or run locally to protect sensitive information, though data still passes to whichever large language model backend connects to Paper2Agent3

3

. James Zou advises that users with protected health information should exclude sensitive data from the framework3

3

.

Open-Source Release and Future AI-Embodied Knowledge Networks

Paper2Agent is now available as an open-source tool on GitHub, with a live online version demonstrating the framework's capabilities3

3

. The Stanford researchers plan to develop an online platform where paper agents can collaborate and discuss various "agentified" scientific discoveries3

3

. Zou hopes the open-source community will help improve the framework while his team continues development3

3

.

This shift from passive artifacts to AI-embodied knowledge networks could eventually link millions of papers, surfacing discoveries human researchers might otherwise miss while crediting original authors2

2

. The framework is compatible with various AI coding agents, though not all have been tested3

3

. As Zou states, this represents "an opportunity to fundamentally reimagine what knowledge looks like" by converting static records into active embodiments of scientific understanding2

2

.