Paper2Agent converts research papers into runnable AI agents

A new paper in Nature proposes turning research papers themselves into runnable AI agents, using a framework called Paper2Agent that packages each paper's code, data and methods into a Model Context Protocol server a chat assistant can call directly. "We introduce Paper2Agent, an automated framework that converts research papers into AI agents," write Jiacheng Miao, James Zou and co-authors in the abstract.

In a case study on Google DeepMind's AlphaGenome, the authors report that Paper2Agent generated 22 validated tools in about 45 minutes at a compute cost of roughly $14, and reached 98.7 ± 1.3% accuracy on tutorial-derived queries, versus 82.7% for a baseline that hands Claude direct repository access. On a wider set of 100 computational biology papers, the framework produced usable agents for 74 of them, scoring 91.2 ± 1.6% on 300 benchmark questions.

The stated ambition is bigger than a benchmark. Paper2Agent, the abstract says, "transforms research output from passive artifacts into active systems that can accelerate downstream use, adoption, and discovery," and one of the case studies claims an assembled AI co-scientist "identified new splicing variant associated with ADHD risk." Five researchers we track in our Who's Who directory posted the link, an early signal the paper is being read beyond its computational-biology core.

About a quarter of the tested papers never made it that far, tripped up by missing code, environment problems, or scripts that don't generalize. The authors argue journals should add an "agent availability" clause alongside the data and code requirements they already impose.