A scientific paper can describe an extraordinary discovery and still leave its next reader facing a surprisingly ordinary problem: how to make the discovery useful. The explanation may be beautifully written, the figures persuasive and the results important, yet the actual method often lives somewhere else. A researcher must find a code repository, install unfamiliar software, locate the data, understand which settings matter and work out how the pieces fit together. Sometimes the instructions no longer work. Sometimes the necessary files are missing. Between reading about a scientific method and successfully using it, there can be days of work—and occasionally a barrier that never comes down.
This separation between knowledge and action has existed in different forms for centuries. Since the earliest scientific journals, researchers have relied on written accounts to share what they observed and how they reached their conclusions. Digital publishing made those accounts easier to find, and public repositories allowed authors to release code and datasets alongside them. Interactive notebooks brought executable calculations closer to the words that explained them. Even so, much of the burden of translating a publication into a working process remained with the person trying to use it. The paper could tell you what was done, but it could not usually do it with you.
Researchers at Stanford approached that problem from a different direction. Instead of asking only how artificial intelligence could summarize a scientific article, they asked whether AI could make the article's underlying methods available as working capabilities. The result, Paper2Agent, was developed by Jiacheng Miao, James Zou and their colleagues and published in Nature. Their proposal was simple enough to sound almost fanciful: transform a research paper into an AI agent that can explain the work, operate the associated tools, analyze new data and cooperate with tools derived from other papers. A method would no longer have to wait for every new reader to assemble it from scratch.
The phrase “a paper becomes an agent” needs an important explanation. The PDF does not wake up, acquire a mind or become the original scientist in digital form. A general-purpose AI system remains responsible for understanding the user's request and deciding which steps to take. What changes is the collection of resources and verified actions that the system can access. Paper2Agent constructs an interface around the published research, allowing an AI assistant to call specific functions drawn from that work instead of improvising the science from prose alone. It is closer to turning a printed laboratory manual into an organized collection of working instruments than to giving a journal article a personality.
To understand why that matters, consider the difference between asking an AI what a paper says and asking it to perform a procedure described in the paper. A language model can read a passage about the effect of a genetic variant and offer a plausible explanation. But if the question requires running a specialized prediction model on a new DNA sequence, eloquence is not enough. The answer depends on the correct software, inputs, parameters, reference data and calculations. A fluent account of a procedure can sound convincing even when no procedure has been executed. Paper2Agent is designed to narrow precisely that gap between talking about an experiment and running its computational method.
The system begins by gathering what the authors have made available: the manuscript, supplementary files, datasets, repositories and examples of how the research code is intended to run. A group of specialized AI workers then examines these materials. One helps set up an isolated software environment, another identifies tutorials and useful components, and others attempt to execute the examples and turn their logic into reusable functions. These are practical, often tedious tasks that researchers recognize immediately. Scientific software can depend on a particular version of a library, require a credential for an external model or expect data to be arranged in a very particular way. Getting the original example to work is a necessary first step toward applying it to something new.
Successful examples are not simply copied into a chatbot. Paper2Agent tries to extract a general procedure from them: what information goes in, what operation is performed and what comes out. A script that was written for one particular file might become a function that accepts many compatible files. The system runs checks against the source examples, comparing results and figures with reference outputs, and it revises tools that fail. Tools that repeatedly cannot pass the checks are left out of the executable interface. This is a crucial safeguard. It does not prove that the original science is correct, but it reduces the chance that the agent will claim to possess a function it cannot reliably execute.
All of this is packaged using the Model Context Protocol, usually called MCP. The technical name is less important than the role it plays. MCP is a common way for AI assistants to discover and use external tools and information without needing a completely different custom connection for each source. In a Paper2Agent package, one part offers resources such as the paper's text, tables and metadata; another exposes callable tools that run the research methods; and a third can provide guided workflows so that a complicated series of steps happens in the right order. A scientist can connect a compatible AI assistant to these components and ask a question in ordinary language while the relevant software operates underneath.
There is an important distinction between these layers. Resources help an agent locate and interpret information. Tools allow it to execute a computational action. Workflow instructions help it combine actions correctly, including steps the human user may not know to request separately. A paper whose code is unavailable might still become a useful, searchable research resource, even though it cannot provide functioning analytical tools. A paper with accessible, well-maintained software can offer considerably more. This is why the Stanford team's work is broader than producing an unusually informed question-and-answer bot: it attempts to preserve the practical know-how of research as something another system can actually use.
The researchers described the result as a “virtual corresponding author,” a phrase that captures both the attraction and the limitation of the idea. It can give a newcomer a guided introduction to unfamiliar methods and demonstrate what an author has publicly documented. Yet no agent can reliably explain the private decisions, failed experiments or unrecorded judgments that never entered the manuscript or its associated files. It cannot promise to represent the author's current beliefs or speak on the author's behalf. The useful achievement is narrower and more concrete: a conversational interface to methods and evidence, with a traceable relationship to their original sources.
One of the clearest demonstrations involved AlphaGenome, an AI model developed to predict how changes in DNA may influence biological processes such as gene expression and RNA splicing. Working directly with a model like this can require specialized programming knowledge and careful control over the kind of prediction being requested. Paper2Agent turned AlphaGenome's software interface into 22 validated tools. In the researchers' published account, that conversion took roughly 45 minutes and cost about 14 US dollars in computing and model usage for this particular example. Once built, the tools could be reused; the cost and time were not a promise that every scientific paper could be converted so quickly or so cheaply.
Imagine a biologist with a genetic variant that has appeared in a study of disease risk. Instead of writing an entire analysis program, the biologist could ask the connected agent to examine the variant in a relevant tissue, compare several predicted biological effects and produce visualizations. The agent would select and run the appropriate AlphaGenome tools and return the outputs with an explanation. The scientist would still need to judge the biological meaning and the quality of the predictions. A computational score can help narrow a search, but it does not by itself establish that a gene causes a disease or that a treatment will work. What changes is the amount of technical assembly required before the scientific judgment can begin.
A particularly revealing AlphaGenome exercise concerned a DNA variant associated with LDL cholesterol. When the paper agent revisited the evidence, it highlighted SORT1, whereas the earlier published analysis had emphasized nearby genes including CELSR2 and PSRC1. Independent genetic-expression evidence discussed by the authors supported several possible interpretations. The interesting result was not that the agent had decisively corrected the original scientists. Rather, the method had become easy enough to rerun and interrogate that an alternative explanation could be examined without rebuilding the entire pipeline. A static conclusion had become something researchers could question with the published tool itself.
The same idea was tested beyond genome prediction. TISSUE is a research method for analyzing spatial transcriptomics, which studies patterns of gene activity while retaining information about where cells sit in a tissue. Its calculations also deal with uncertainty, an essential detail when researchers are making predictions about measurements they cannot observe directly. The Paper2Agent version could help users understand which datasets were needed, find associated resources and run an analysis that produced prediction intervals. In the team's demonstrations, its outputs matched reference work performed by human experts. The contribution was not merely saving typing; it made a demanding method accessible through a guided process without discarding the uncertainty built into the underlying science.
Another example used Scanpy, widely employed for studying gene expression in individual cells. A reliable Scanpy analysis normally involves many linked decisions: filtering poor-quality cells, normalizing measurements, choosing informative features, reducing the number of dimensions, identifying related cells and grouping them into clusters. Doing those operations in the wrong order or with unsuitable settings can distort the result. Paper2Agent exposed relevant functions as reusable tools and used MCP workflow instructions to guide their sequence. It was a demonstration of something easily missed in discussion of AI agents: sometimes intelligence is not inventing a new procedure but carrying out a well-established one correctly, every time the data requires it.
Up to this point, each paper agent might seem like a helpful doorway into a single publication. The more ambitious possibility appears when an AI assistant can connect to several of them. One study may offer a prediction model, another an experimental dataset and a third a method for testing whether two patterns agree. The publications need not have been written by the same team or designed to work together. If their methods and data can be translated into compatible interfaces, a coordinating agent can draw on all of them in one investigation. This is the moment at which a library starts to look less like a collection of finished reports and more like a collection of scientific instruments that can be arranged into new experiments.
The Stanford researchers explored this in an investigation of psoriasis, an immune-related condition that affects the skin. They assembled three paper-derived agents with different strengths. AlphaGenome offered predictions about how a genetic variant might influence nearby genes. A second source contained results from a study that perturbed regulatory DNA in immune cells using a combination of reporter assays and CRISPR interference. The third offered Perturb-seq data, which records changes in gene activity after particular genes are experimentally disrupted. Put together, these sources allowed a question to be approached from several directions: a genetic prediction, an intervention affecting DNA regulation and an intervention affecting a candidate gene.
The investigation focused on a psoriasis-associated variant, rs887314. AlphaGenome placed GPR137 high among the possible affected genes in CD4-positive T cells, a type of immune cell involved in inflammatory responses. But a high predicted score alone was not enough to show which gene mattered. The AI co-scientist examined the other papers' data and proposed ten approaches for further analysis. A human researcher selected a comparison of gene-expression signatures: if changing the regulatory DNA and knocking down GPR137 produced similar downstream effects, that would strengthen the proposed link. This human selection is worth noticing. The demonstration combined autonomous computation with a consequential scientific judgment made by a person.
The comparison produced encouraging results. Among the leading candidates with suitable data, GPR137 showed a significant match to the regulatory perturbation pattern when the immune cells were stimulated, but not when they were at rest. The reported correlations were approximately 0.61 and 0.63 under two stimulation conditions. This added a more specific possibility to the original prediction: the gene's role at that genetic location may depend on the activation state of the cell. The evidence was drawn from previously generated experimental datasets; Paper2Agent did not perform a new physical CRISPR experiment. The combination of sources supported a biological interpretation, but independent experimental follow-up would still be needed before treating the causal account as settled.
A second demonstration connected AlphaGenome to a genetic association study of attention-deficit/hyperactivity disorder. An AI co-scientist explored 209 candidate variants and highlighted rs1626703, predicting an effect on the splicing and expression of a gene called MPHOSPH9 in particular brain cells. The system could then extend similar analyses across additional disease-associated regions. This is the kind of connection that could be easy to miss when publications and datasets are examined separately. It is also a good example of how carefully the word “discovery” should be used. The resulting mechanism was a computational hypothesis. It did not establish an explanation for ADHD in patients, nor did it provide a clinical test or treatment.
Both examples suggest a change in the scale at which researchers can combine published work. Human scientific collaboration often depends on knowing whom to contact, understanding another laboratory's software and finding time to adapt unfamiliar methods. Paper agents may lower some of those barriers. A geneticist could bring a statistical tool from another field into a familiar workflow, or a laboratory studying one disease could examine whether a model built for another question has something useful to say about its data. The possibility is not that an AI will automatically know which ideas deserve trust. It is that more combinations become practical to investigate, and that their intermediate steps can be preserved and examined.
The team also tested Paper2Agent at a scale larger than a few carefully selected successes. Out of 100 computational-biology papers processed without manual repairs, 74 were successfully converted into working agents. For those papers, the system proposed 599 executable tools, and 593 passed its automated validation. On 300 questions based on the source tutorials, Paper2Agent achieved an average accuracy of about 91 percent using Claude Sonnet 4, compared with roughly 80 percent for an assistant using the same model with direct access to the repositories. The paper also reported a stronger-model direct-repository comparison of about 86 percent. In this evaluation, giving the model well-prepared, tested tools mattered more than simply handing it a pile of source code.
Efficiency improved as well. The researchers reported an average query cost of about 20 US cents and a response time of 1.6 minutes for Paper2Agent, compared with 38 cents and 4.3 minutes for their direct paper-and-repository baseline. These figures belong to a particular study with particular models, tasks and pricing; they should not be treated as fixed costs for future users. More importantly, the scores measure performance on benchmarked questions and executions. They do not mean that 91 percent of all scientific claims in the converted papers were verified, or that the tool can be trusted on every new dataset without review.
The results extended beyond computational biology. In 42 execution-based tasks drawn from ten papers in fields including artificial intelligence, statistics, econometrics and game theory, the study reported roughly 98 percent accuracy across repeated runs. The team also evaluated papers centered on datasets and scientific findings rather than readily executable software. For 26 such papers, the structured resource layer answered synthesis questions with about 89 percent accuracy, compared with 82 percent for a browser-based AI baseline. Those are promising demonstrations of portability, although the tested collections were still small relative to the enormous variety of scientific literature. A method that works well on documented computational research may face very different obstacles in chemistry, field ecology or clinical practice.
The 26 unsuccessful conversions among the initial hundred papers may be as instructive as the successes. Missing executable code, unavailable datasets or model files, dependency failures and scripts that did not transfer beyond their original examples prevented the pipeline from producing working agents. These are not all failures of artificial intelligence. They expose long-standing weaknesses in scientific software and the difficulty of reconstructing an experiment from what gets published. Paper2Agent may therefore become a useful stress test for computational reproducibility: not a final judgment on a paper's scientific merit, but a practical indication of whether another system can assemble and run what its authors described.
That distinction between computational reproducibility and scientific truth is essential. An agent can rerun a method perfectly and reproduce a mistaken result if the original assumptions were wrong. It can deliver a precise answer to a poorly chosen question. It can combine two datasets whose differences make their comparison misleading, or mistake a prediction for experimental evidence. Tests against reference code improve faithfulness to the method, not the validity of every interpretation a researcher might draw from it. The system may make the mechanics of science faster, but it cannot abolish uncertainty, statistical judgment or the need for independent replication.
There are also less visible challenges. Some studies use restricted patient information, licensed software or datasets that cannot safely be distributed through a public tool server. Research code can contain vulnerabilities, and agents capable of executing it need tightly controlled environments, limited permissions and records of what they did. A malicious instruction hidden inside an untrusted repository is a different problem from an ordinary scientific mistake; it could redirect an agent away from its authorized task. When several paper agents contribute to a result, provenance becomes equally important: the final analysis should show which paper supplied a method, which dataset supplied evidence, which tool version ran and where human decisions entered the workflow.
Attribution matters for a deeper reason, too. The paper agent did not design the original experiment, recruit its participants, maintain its instruments or make the difficult judgments behind the published method. Those achievements belong to researchers and research communities. If a new finding emerges from five connected paper agents, credit should not disappear into the name of the AI system that coordinated them. The developers of the methods, the creators of the datasets and the scientists who framed and evaluated the new investigation all remain part of the discovery. A future built around executable research will need better ways of preserving that intellectual history, not fewer.
The most promising change may eventually reach the format of scientific publishing itself. Today, authors commonly state where their data and code can be found. The Paper2Agent researchers suggest that journals might one day add something like an “agent availability” statement: a clear way to access the working methods of a paper through a conversational interface. That would be an evolution of earlier efforts such as executable notebooks and open repositories, not a sudden replacement for them. It could also alter what good documentation looks like. A paper that exposes its assumptions, software dependencies and reusable workflows clearly is more likely to remain useful to both people and machines.
It is tempting to imagine millions of these paper agents quietly searching for connections, matching methods with datasets and proposing experiments that no one thought to combine. Stanford's team has described that possibility, and the early demonstrations explain why it is worth investigating. But expanding from tens or hundreds of paper agents to an entire interacting scientific literature would create new problems of quality control, coordination and trust. Incorrect methods could be reused at scale as easily as correct ones. An impressive computational hypothesis might circulate long before any laboratory tested it. Larger networks would need stronger systems for evaluating evidence, managing updates and knowing when to stop and ask a human scientist.
None of this diminishes the achievement already demonstrated. There is a meaningful difference between an AI that has read about a technique and one that can invoke the technique through documented, tested tools. Paper2Agent begins to make that difference available beyond the research groups that originally wrote the code. It offers a way for published knowledge to travel not only as explanations but as reusable methods, complete with the materials, instructions and checks that make their use possible. The significance lies less in the image of papers becoming living beings than in a quieter transformation: scientific ideas becoming easier to examine, reproduce, challenge and combine.
For centuries, a scientific paper has asked its reader to do something demanding: understand the argument, trust only what the evidence supports and carry the work forward. That responsibility has not changed. What may change is the distance between understanding a method and being able to try it. A library of papers will always remain a record of human inquiry. With carefully built paper agents, it may also become a place where that inquiry can be continued—one accountable, testable calculation at a time.
Sources
Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard and James Zou, “Reimagining research papers as interactive and reliable AI agents,” Nature. https://www.nature.com/articles/s41586-026-11044-y
Stanford Medicine, “Manuscripts-turned AI agents can now ‘talk’ to each other, make new discoveries.” https://med.stanford.edu/news/all-news/2026/09/ai-agents-talk.html
Paper2Agent research team, official code repository, examples and implementation notes. https://github.com/jmiao24/Paper2Agent
Nature, “AI tool turns any paper into an ‘agent’ that can collaborate and answer complex queries.” https://www.nature.com/articles/d41586-026-02899-2