A Stanford team has presented Paper2Agent, a system designed to convert a scientific publication and its supporting materials into a runnable AI agent. Reported in Nature, the approach ingests a paper’s text, code, data and methodological details, then packages those materials as an interactive agent and Model Context Protocol (MCP) server that can be queried and used through compatible AI tools.[1]
The significance is less about producing another literature chatbot than about proposing a different unit of scientific publication: a paper that can be inspected, executed and extended as software. In a demonstration built around AlphaGenome, the researchers created an agent in roughly 45 minutes for about $14. The agent answered genetics questions with near-perfect accuracy in the reported evaluation and produced an alternative hypothesis concerning a cholesterol-related gene variant.[1] Those results are promising, but they also underline the system’s central limitation: an agent can only operationalize what its underlying paper, code, data and documentation make available.
By the numbers
- About 45 minutes: Reported time to create the AlphaGenome agent.
- About $14: Reported cost of creating that agent.
From static paper to research interface
A conventional paper is optimized primarily for human reading. It describes a result, summarizes methods and often points readers to repositories, supplementary files and datasets. Reusing the work can require a separate and time-consuming reconstruction process: locating the right software version, installing dependencies, interpreting preprocessing decisions, finding input files and determining which parts of a workflow are essential.
Paper2Agent attempts to turn that reconstruction work into a software-generation task. The system uses the materials associated with a publication—its narrative explanation, source code, data and methods—to assemble an agent with access to the study’s relevant capabilities. Rather than merely retrieving passages from the article, the resulting agent can expose tools for working with the research artifacts and can explain or test aspects of the reported work.
The MCP component is important. MCP is a protocol for connecting AI systems to external tools and data sources through a standardized interface. Packaging a publication as an MCP server means the research workflow can potentially be used by different MCP-compatible clients rather than being trapped inside a single chatbot interface. In practical terms, an analyst could ask questions in natural language while the agent invokes functions based on the paper’s code and data, subject to whatever access controls and computational environment the publication provides.
That architecture makes Paper2Agent meaningfully different from a paper-summary assistant. A summarizer can restate what authors wrote. A runnable agent can, in principle, execute defined operations, expose parameters, compare inputs and return results grounded in the supplied research package. The distinction matters most when the question is not “what did this paper claim?” but “can this result be reproduced, perturbed or applied to this new case?”
The AlphaGenome demonstration
The Stanford team’s AlphaGenome example provides the clearest indication of the intended workflow. According to the Nature report, the system produced an AlphaGenome-focused agent in about 45 minutes at a cost of roughly $14.[1] The agent answered genetics questions at near-perfect accuracy in the authors’ reported assessment and generated an alternative hypothesis about a cholesterol-related gene variant.
The hypothesis-generation result deserves careful interpretation. It does not establish that the agent discovered a validated biological mechanism, nor does it mean an AI system can independently settle a genetics question. It shows a more bounded but still useful capability: when a model can access a research system’s methods and computational machinery, it may combine the published framework with a new line of inquiry and articulate a testable alternative. That is closer to a research collaborator’s workflow than to a static question-answering system.
For genetics, where models often transform sequence information into predictions about functional effects, an interactive research wrapper could lower the barrier for domain scientists who need to interrogate a model without rebuilding its full software environment. A researcher might ask which assumptions affect a variant prediction, request a comparison across candidate variants or inspect how a reported result changes under an explicitly defined condition. The scientific value would depend on whether the agent preserves the underlying model’s provenance, faithfully exposes its limits and makes each computational step auditable.
Why executable publishing could matter
The period between publication and practical reuse is an expensive source of friction in science. A result may be public while its implementation remains difficult to run. Even when code is available, it may lack dependency specifications, stable data access, complete configuration files or documentation of the informal decisions made during analysis. The reproducibility literature has documented how weak incentives and the practical burden of research software preparation can discourage code from being made reusable.[2]
Paper2Agent points toward a publication model in which readers do not begin from a PDF alone. They begin with an accessible, constrained interface to the paper’s own computational methods. If such interfaces prove dependable, they could help labs evaluate whether a method fits their question before investing days or weeks in setup. They could also make peer review more concrete: reviewers and subsequent researchers could probe stated claims, test supported inputs and identify missing information in an executable workflow.
There are potential benefits beyond replication. A well-designed agent can serve as a durable onboarding layer for a lab’s work, explaining expected inputs, identifying supported uses and connecting methodological prose to the specific code path that implements it. It can make a complicated computational publication more usable for collaborators, educators and researchers in adjacent fields who have expertise in the science but not in every implementation detail.
For publishers, repositories and research-software platforms, the idea creates a possible product and infrastructure category: executable supplements with standardized tool interfaces, provenance records, permissions and evaluation suites. That is a projection, not a measured market outcome. Adoption would depend on incentives for authors, durable hosting, compute costs, licensing, security and discipline-specific standards for validation.
Execution does not eliminate scientific uncertainty
Making a paper runnable does not automatically make it reproducible, correct or broadly applicable. Paper2Agent inherits the quality of the material it receives. If the paper omits crucial preprocessing details, if the code cannot be executed reliably, if datasets are incomplete or restricted, or if the original method contains a conceptual error, an agent cannot repair those gaps simply by presenting a polished conversational interface.
The system also introduces new evaluation questions. An agent must be tested not only on whether it answers factual questions about a paper, but on whether it calls the correct tool, uses the right data version, handles invalid requests safely and distinguishes results it computed from explanations inferred by a language model. Reproducibility requires provenance: users need to know what code version ran, what parameters were selected, which data were accessed and whether an output matches the authors’ original environment.
Natural-language interaction can create a false sense of authority if the interface does not clearly separate documented facts, executable results and speculative suggestions. The AlphaGenome agent’s alternative hypothesis is valuable precisely because it is a hypothesis to test, not a conclusion to accept. In biomedical work especially, agents should surface uncertainty, cite the source operation behind an answer and preserve a reviewable record of analyses.
Security and governance will also matter. Research agents that can execute code or access datasets need sandboxing, permission boundaries and controls around sensitive data. Some of the most consequential research datasets, including clinical and proprietary collections, cannot simply be exposed through a general-purpose conversational endpoint. A mature executable-publication ecosystem will need to support local deployment, authentication, access logging and clear restrictions on data use.
What comes next
The strongest test for Paper2Agent will be whether it works outside polished demonstrations and across the uneven reality of published research software. Useful benchmarks would measure setup success, reproducibility of headline results, accuracy of tool use, quality of provenance records and the time saved by independent users. Comparisons should include not only chat-based assistants, but the existing baseline: an experienced researcher attempting to reproduce the work from the paper, repository and supplementary materials.
The approach may be most effective in fields with computationally explicit methods, reusable datasets and established evaluation tasks, including genomics, machine learning, materials science and parts of computational chemistry. Its impact could be greatest where a method is valuable but difficult to install or interpret. Conversely, papers built around unavailable data, bespoke laboratory procedures or poorly maintained code will remain hard to operationalize.
The broader shift is toward treating research outputs as interoperable systems rather than documents alone. Paper2Agent does not solve the longstanding incentives behind incomplete code and documentation. It does, however, offer a practical reason to improve them: better structured papers and research artifacts can become directly useful interfaces for scrutiny and follow-on work.
Editor’s Take
I think the compelling part of Paper2Agent is the packaging, not the conversational layer. Most teams do not need an AI to paraphrase a paper; they need a reliable path from a claimed result to a runnable method. If an agent can expose the actual workflow, preserve inputs and parameters, and make it easier to test a result on a new question, it can remove a very real tax on research and product development.
The next thing to watch is whether these agents retain a complete execution trail and work on messy, independently authored papers. The $14, 45-minute AlphaGenome example is an encouraging proof point, but it is not yet evidence that arbitrary publications can be converted into dependable scientific infrastructure. The winners will be systems that make uncertainty and provenance visible—not systems that make a paper sound more certain than its artifacts justify.
