Stanford University introduced the Paper2Agent framework for turning scientific papers into AI agents

A team at Stanford University led by James Zou published the open-source framework Paper2Agent in Nature on 16 September 2026. It automatically extracts the workflows described in a supplied academic paper and accompanying code, data or other materials and turns them into a tested, executable AI agent. Unlike tools such as Gemini Notebook (formerly NotebookLM), which only allow conversation about the contents of a document, Paper2Agent agents, according to the authors, actually run the methods described in the paper and can combine them with tools derived from other papers. The framework was tested across disciplines including statistics, econometrics and astrophysics, but the main demonstrations focused on computational biology.

The demonstration used the AlphaGenome model, which predicts the effect of DNA mutations on gene regulation, supplemented by the previously released AlphaGenome Atlas catalogue containing predictions for all the billions of possible single-letter changes in the human genome. After receiving the documentation and code, Paper2Agent created 22 tools covering various AlphaGenome functions without human intervention in approximately 45 minutes on a standard laptop, with total computing costs under 15 USD. Among other things, the tools predict the impact of a specific DNA change on gene activity, compare this effect across tissues or analyse multiple variants at once. According to the authors, all tools passed automatic verification by a testing agent that diagnosed and fixed problems when a test failed, with up to six attempts per function; tools that could not be fixed were discarded.

The verified tools were packaged into a Model Context Protocol server and connected to Claude Code, but according to the authors, the setup also works with another compatible AI assistant. According to the analysis by the authors, the resulting AlphaGenome agent outperformed both the standard Claude model with the same AlphaGenome code and the specialised tool Biomni. The team then connected agents created from two other papers (on inherited DNA variants associated with autoimmune diseases and on systematic gene silencing in immune cells) to the AlphaGenome agent. When given a query about the genetic basis of psoriasis, the agents jointly identified the gene GPR137 as a likely causal factor and proposed ten ways to verify it; a human researcher selected one of the options, and subsequent analysis confirmed a similarity to the effect of a variant associated with psoriasis.

Details, including further development mentioned in the source, can be found in the source article.