Executive Summary

Key Figures

Sources: Nature's news story, Stanford Report, and the authors' preprint. Where each figure comes from is linked in the body.

James Zou, the professor who led the work, set out in Stanford Report how people have handled knowledge so far. Long ago it was carved into stone, now it is printed as letters on paper, and paper is not much of an improvement on stone. Either way it is a passive object that opens only when a reader comes to it. The proposal that followed was to turn a static record into a form where the knowledge moves on its own.

The method is built around a standard called MCP. Zou compared it to a filing cabinet holding a paper's PDF sorted into a shape an agent can find and use. Introduction, methods, results and conclusion each go into a different drawer, and the drawers sit inside one box called the paper.

Three kinds of thing go into that cabinet. Tools, which wrap the paper's methods as runnable functions. Resources, which hold the manuscript and the code and the supplementary material in a standard form. And prompts, which record what order a multi-step procedure runs in. That last one stands out. The server for the single-cell analysis package carries the whole sequence from quality control through normalization, feature selection, dimensionality reduction, clustering and cell-type assignment, and the authors state that nobody wrote this order out by hand; it was pulled from the paper and the codebase.

Several AIs share the job of filling the cabinet. One finds and downloads the code repository, one stands up the runtime, one pulls the core capability out of the tutorials and wraps it as a tool. What comes next defines the character of this tool. For every tool it builds it writes tests, runs them, reads the cause of any failure and applies a fix. Functions that keep failing after the fixes get dropped from the list. The design keeps only what actually runs the way the paper says it does.

Where the raw material comes from reveals the premise of the whole procedure. What the tool wraps is not the paper's method code as a whole but the tutorials sitting in the repository. The AI in charge scans the repository to separate genuine tutorials from other files, then runs the chosen tutorials end to end on their example data, writing down inputs, outputs, figures and run conditions. At this step it also records assumptions that were left implicit, the preprint says. Testing uses only the examples the tutorial itself carries. A tool that passes is locked in that state, and each tool carries a pointer back to the place in the original code it came from. The design is aimed at stopping an AI from inventing plausible code and returning a wrong result.

The finished server goes up somewhere like Hugging Face, and a user only attaches the address to whatever chat agent they already use. Nature's news story called the result a virtual corresponding author. Ask it something and it answers, it applies its own paper's method to new data, and it talks directly to agents built from other papers. By Stanford Report's account, more than 100 paper agents have been built so far. The code and instructions are public, and the servers built from AlphaGenome and the single-cell analysis package work as soon as the address is attached.

Attempts of this kind are not new. The preprint marks the earlier line itself: a publisher's competition for executable papers, publishing that put Jupyter notebooks into the text, and Papers with Code linking papers to their repositories. Reproducibility rose each time and a barrier stayed, in the authors' assessment. Finding code got easier, and installing it and running it stayed with the reader.

The first paper the team attached this to was DeepMind's AlphaGenome. The model predicts what a single changed DNA base does to gene expression, and using it means setting up an environment, importing several modules, building a client with an API key and matching the input format. For a biologist working alone the bar is fairly high.

Twenty-two tools came out of this paper without human intervention. Nature's news story puts the time at about 45 minutes and the computing cost at $14. The preprint the authors released last year, though, gives around three hours on a personal laptop for the same 22, and the 45 minutes there belongs to the single-cell analysis package, where seven tools were built. The final version's text is paywalled, so this article could not confirm which way the two were settled.

On accuracy the preprint's numbers are the specific ones. Given 15 questions drawn from the AlphaGenome tutorials and 15 fresh questions that were not in them, this agent got both sets right. Claude, a general-purpose coding agent handed the same repository whole, got 9 and 12. Biomni, a biomedical agent that draws on dozens of databases, got 6 and 9. Nature's news story summarized the result as near-perfect accuracy.

The agent also returned a different answer from the original paper. Asked which gene is responsible for a base variant linked to bad cholesterol levels, the agent named SORT1 as the strongest candidate. The genes the AlphaGenome paper emphasized were CELSR2 and PSRC1. The agent gave two grounds: a large predicted effect on SORT1's expression in liver tissue, and the fact that the protein SORT1 makes is directly involved in secreting bad cholesterol. On the scores written into the preprint, though, the value AlphaGenome itself assigned ran slightly higher for CELSR2 and PSRC1. When the authors went and checked public data, all three genes showed a significant signal in liver tissue. Pinning a single cause at a spot where several genes line up is hard to begin with, the preprint concludes. Zou took this as a strength instead. Published conclusions can be looked at again without designing a new experiment.

The researchers also tried two papers at once. The AlphaGenome paper and an ADHD genome-wide association study were each made into a server and attached together to one AI, and that AI opened by putting forward several hypotheses worth checking. That regulatory activity shifts in particular brain cells. That a causal variant can be narrowed out of the candidates statistics had already bracketed. That transcription factor binding breaks at the FOXP gene locus. It was a person who picked the second one and told it to run. From there the AI drew up its own analysis plan and swept 39 gene loci. It took under two hours. One of 209 candidate variants, rs1626703, came up as causal, and a path was proposed in which the variant makes an exon of the MPHOSPH9 gene get included more often and raises expression in excitatory neurons. Zou said the link had not been reported before. It is a candidate the computation pointed at, not a discovery confirmed by experiment.

A case going well does not mean the thing works on any paper at all. The researchers gathered 100 computational biology papers and attempted the conversion, and working agents came out of 74. The 26 failures sort into three groups. The code was incomplete, or documentation was missing, or the software environment could not be revived. The numbers for this trial are printed in the final version's text and that text is paywalled, so this article confirmed them through the coverage.

The authors wrote the limit into the preprint themselves. If the original codebase is incomplete, poorly documented or carrying unresolved errors, their sentence runs, the tool cannot reliably expose it as a working tool. Put the other way around, swapping in a better model does not make missing code appear.

3.1What Automation Repairs and What It Cannot

Plenty of breakage can be repaired. The AI on the job diagnoses and patches common problems like broken dependencies, mismatched file paths and outdated API calls. The loop of running tests, reading failures and fixing again does that work. The line it does not cross lies in front of all that. Code and data that were never released leave nothing to restore.

To the authors, how easily a paper turns into an agent can serve as a practical measure of that research's reproducibility and rigor. The prediction that follows is more direct. Just as the field came to expect data and code to be released as a matter of course, expecting contributions to be structured so that they translate easily into agents is the next step, the authors wrote.

The forecast extends to a new box in the paper itself. Much as many journals now ask where data and code can be obtained, there would be a place to state whether the research has been embodied as an agent and where it is attached. The preprint already carries that box. The address hosting the AlphaGenome agent is printed at the end of it.

So far the work has covered methodological papers. Papers that put forward algorithms, models or computational workflows, where what to carry across is clearest, is how the authors explain the choice. Widening to papers that release a data resource or report a discovery is left as future work, and there they expect the agent's job to shift from computation toward interpretation and curation.

3.2What the Manuscript Never Holds

Even a paper that converts well has an empty box. Stanford Report pointed out that a manuscript does not carry the failed experiments or the judgment behind an experimental setup. So the human author has to fill that context in by talking to the agent built from their own paper. The agent asks the author about the paper, the author answers, and the empty box fills in over that exchange.

This is where the tool draws its boundary. Automation can only carry across what was recorded. Which method was tried first and abandoned, why this threshold was picked, which data was dropped and for what reason: mostly none of it reaches the manuscript. Nothing that went unwritten is available to any model.

One distinction stays with us whenever we talk about AI-Ready Data. Having material and having it ready to use are two different states. This result attaches a scale to that distinction. If the same tool run the same way carried 74 papers and stalled on 26, the difference came from the state of the material each lab left behind rather than from the tool. Narrowed further, it came from whether anyone wrote down examples somebody else could follow. That is what this tool takes as its raw material: the tutorials in the repository.

This blog covered a similar scene back in June. In an experiment where AI agents ran released paper code themselves, the success rate came to 54.1%, and what blocked it was the runtime, the dependencies and data alignment. This study confirms the same wall from the other side. Released and runnable are different things, and runnable and explainable are different again.

Moved onto company work, the handover document stands in the same place. When a dataset goes to another team or an outside supplier, the schema, the files and the labeling rules generally travel with it. Why the cut was made at this threshold, which label definition was tried and dropped and why, which samples were removed for what reason: those stay in one person's head. When that person changes teams the data is unchanged and the grounds for the judgment are gone. The next person repeats the same trial and error from the beginning.

Four checks give a rough sense of which side our own material is on today.

• Can somebody who just arrived stand up the environment from the documentation alone and run the pipeline end to end? If they cannot, that pipeline is not an asset yet; it is one person's memory.
• Is it written down anywhere what was tried and dropped before the current rules were settled on? A record with not one failed attempt in it is usually a record that stopped short.
• When a case arrives that the label definition does not cover cleanly, are there adjudicated examples to work from? Ten ambiguous cases beat ten pages of taxonomy.
• If the person in charge is away tomorrow, does the documentation still say what this data was meant to help judge?

Reproduction will certainly get easier once papers answer for themselves. What gets easier first, though, is research that was already left in good order. The better the tools get, the further apart well-kept material and the other kind end up.

Thank you for reading this far. The figures and sentences this article cites can be checked by anyone in Nature's news story, Stanford Report, and the preprint the authors released. We would be glad to hear whether the last data handover document you sent listed the attempts that failed, and if it did not, what stood in the way.

Academic Papers

• 1.Miao, J. & Zou, J. (2026). "Reimagining research papers as interactive and reliable AI agents." Nature.
• 2.Miao, J. & Zou, J. (2025). "Paper2Agent: Reimagining Research Papers As Interactive and Reliable AI Agents." arXiv:2509.06917.

News Coverage

• 3.Glickman, K. (2026). "AI tool turns any paper into an 'agent' that reproduces its own results." Nature News.
• 4.Armitage, H. (2026). "AI agents built from scientific papers surface new discoveries." Stanford Report.
• 5.Stanford Medicine. (2026). "Manuscripts-turned AI agents can now 'talk' to each other, Stanford Medicine-led study shows."

Tools & Code

• 6.Miao, J. et al. "jmiao24/Paper2Agent." GitHub.