Stanford's Paper2Agent turns papers into working AI agents

Stanford researchers built a working AI agent for a genetics paper in about 45 minutes at a compute cost of US$14.

The tool, Paper2Agent, is described in Nature. It reads a paper's main text, code and datasets, deposits them on an MCP server, and then a team of AI agents autonomously builds tools that apply the paper's methods to new data. A scientist connects to the server using the large language model of their choice and queries the paper in plain conversational language, or hands it fresh data to run.

The team, led by James Zou, a computer scientist at Stanford University, tested the pipeline on the AlphaGenome paper, the AI model that predicts properties of DNA sequences. The resulting agent answered genetics questions with near-perfect accuracy.

The technology, Zou says, "can help us to reimagine what knowledge looks like in the future."

Three of the researchers we follow in our Who's Who directory posted the piece the day it went up.

The Nature writeup rests on that one worked example. It does not publish a success rate across other papers, does not say who is responsible for maintaining an agent when the underlying methods are later corrected, and does not describe how a reader distinguishes the agent's answers from a confident confabulation over the paper it wraps.