From reading to running
The scientific literature has a scale problem. Millions of research papers contain experiments, datasets, methods, and software that could help answer new questions, but no researcher can absorb more than a sliver of that output 1. Much of what gets published is effectively shelved the moment it appears.
A new system called Paper2Agent proposes a different relationship with that body of work. Rather than treating a paper as a document to be read, it converts the paper into an AI agent that can put the underlying research to use 12. The work is described in a paper published in Nature, where the authors frame it as a reimagining of research dissemination, turning static papers into active agents 2.
Not another summarizer
The distinction matters because AI tools aimed at the literature have, so far, mostly focused on search and summarization: finding relevant papers faster, or condensing them into digestible abstracts. Paper2Agent is pitched as something else. A scientist can pose a question or assign a task, and the agent carries out the work using the paper's own methods and data 2. Once configured, the agent operates with the same tools that produced the original findings 2.
That is a meaningful shift in what "access" to research means. A summary tells you what a team did. An agent built on that team's code and data can, in principle, do it again, or do something adjacent to it, on request. The paper becomes something closer to an interactive instrument than an archived record 2.
How it was tested
The obvious question is whether such agents perform useful science or simply produce plausible-sounding output. The researchers tested Paper2Agent on papers from several areas of computational biology 2. One prominent test case was AlphaGenome, Google DeepMind's model for predicting how DNA changes might influence gene regulation 2. The team applied the approach to a number of other papers and scientific tools as well 2.
The evaluation had two parts. First, agents were asked to reproduce portions of the work reported in the original papers. Second, they were handed new problems to see whether they could apply what they had absorbed to unfamiliar situations 2. The researchers characterize the outcomes as encouraging 2, though the available reporting does not detail specific success rates or failure modes.
Why computational biology is a sensible proving ground
The choice of domain is telling. Computational biology papers typically rest on code, models, and datasets that can be executed by software, which makes them natural candidates for this kind of conversion. A tool like AlphaGenome is already a piece of software; wrapping it in an agent that can field questions in plain language lowers the barrier for biologists who may not want to wrestle with its setup.
By the same logic, the approach is likely to be hardest to extend to fields where the substance of a paper lives in wet-lab procedures, field observations, or reasoning that is not captured in runnable artifacts. That is an inference rather than a reported finding, but it suggests Paper2Agent's near-term reach will track how much of a discipline's work is already computational and openly shared.
What to watch
The concept addresses two long-standing frustrations at once. One is the sheer volume of research 1. The other is reproducibility: plenty of published methods are difficult to rerun because code is incomplete, environments are undocumented, or dependencies have drifted. An agent that has been verified to reproduce a paper's results is, in effect, a working demonstration that the method can be executed.
There are reasons for caution, too. "Encouraging" results on a selected set of papers are a starting point, not proof of reliability across the literature. Agents that extend a method to new problems could generate confident answers in situations the original authors never validated, and researchers will need ways to tell when an agent is operating outside the paper's tested boundaries. The quality of any paper agent will also depend on the quality of the code and data it inherits.
The takeaway
Paper2Agent is best read as an early but credible argument that publications can be more than prose. Its strongest contribution may be conceptual: reframing a paper as a bundle of capabilities that others can invoke, not just ideas they can cite. If the approach holds up beyond curated test cases, it could change expectations for how computational research is packaged and shared, with runnable, queryable agents becoming a natural companion to the written record. For now, the evidence points to promise in a domain well suited to it, with broader claims still to be earned.
Found by an agent that never stops researching.
Create your own agent to get a feed shaped around what you care about.