Paper2Agent converted 74 of 100 unfiltered computational biology papers into tested agents

The 26 comes out of the validation stage. Paper2Agent runs six stages in order: it locates the paper's repository, builds an isolated software environment, identifies tutorials, executes them, extracts reusable functions, then tests those functions against the paper's original outputs ^[4]. Functions that keep failing after repeated attempts are dropped from the finished server, and the pipeline keeps a record of what ran and what did not ^[6]. A reader sees only the tools that survived that test ^[12].

Tutorials are found and executed before any function is extracted ^[4], so a repository with working code and no tutorial gives the extractor nothing to run. Twenty-six failures out of 100 papers is roughly one paper in four ^[10]. Because each of the 74 successes required code that executed in an automatically built environment, 26 percent puts a ceiling on the share of these papers whose code does not run ^[11]. The paper does not give the full cause breakdown ^[9].

The 100 papers were drawn without filtering for documentation or repository quality ^[7], so 74 conversions ^[8] is a rate over papers as they were actually published. Four conditions have to hold for that rate to travel: a locatable repository, an environment the builder can construct, at least one tutorial that executes end to end, and published outputs precise enough to test a function against ^[4]. The evaluation covered computational biology ^[7]. runtimewire, which reported the work, calls the repository's star count a measure of developer interest and not evidence of scientific adoption ^[14].

"Papers are written for humans," Zou wrote in the announcement thread ^[16]. Zou is a Stanford associate professor of biomedical data science who co-conceived the framework with Miao and supervised the work ^[17]. They name ordinary friction: find the repository, install fragile dependencies, reconstruct undocumented assumptions, work out which inputs produce which outputs ^[22]. Their aim was an agent that runs the work inside a virtual lab and gains experience with the methods ^[18].

The demonstrations were built around AlphaGenome, Scanpy and TISSUE, and those agents reproduced published analyses and handled new queries ^[19]. A second demonstration put three paper agents on a psoriasis-associated locus, connecting AlphaGenome predictions with two independent perturbation datasets, and prioritized GPR137 as the probable causal gene across methods and datasets that had been published separately ^[21]. In a supplementary case, an AlphaGenome agent worked with an ADHD genome-wide association dataset agent to prioritize a variant that may affect the splicing and expression of MPHOSPH9 ^[20]. The paper treats that result as a candidate mechanism requiring experimental validation ^[20].