Scientific papers have long been quiet, static objects. You read them, maybe download the supplementary materials, and try to make sense of the methods from dense paragraphs and equations. A new system called Paper2Agent is aiming to change that. It treats a paper not just as words but as a collection of working tools. When a paper is released with code, methods, and data, Paper2Agent packages all of it into a tool set that an AI assistant can actually operate. That means a scientist can come along later with their own data and ask the assistant to apply the paper’s methods, run the analyses, and answer questions about the results. Stanford computer scientist James Zou calls it “a virtual author of the paper,” available around the clock. The description is apt. Instead of a document that sits in a journal’s archive, the paper becomes an active presence that can explain, compute, and guide. It is as if the original authors left behind a tireless research assistant who knows every line of code and every step of the workflow. Think of the difference between reading a recipe and having a chef standing beside you, ready to adapt it to the ingredients you have on hand. The paper-agent is less a manuscript than a skilled sous-chef. The team describes the approach in Nature on September 16. The broader promise is not just convenience. It is a new way of transmitting scientific knowledge, one where the distinction between “reading about a discovery” and “using a discovery” begins to blur. Rather than spending days trying to reproduce a method from scratch, a scientist can hand new data to an assistant that already knows the paper’s inner workings. That could make computational research more accessible, especially to labs without deep programming expertise. It could also change what it means to publish a study: not merely reporting what you did, but leaving behind something that can keep doing it.
To see whether the concept could work beyond a single lab, Zou and his colleagues tried to convert 100 computational biology papers into working tool sets. They succeeded with 74. The other 26 were defeated by familiar realities: missing code, incomplete data, or software that would not run because of outdated dependencies or broken links. That in itself is a reminder of the gap between how science is presented and how science is actually practiced. But for the successes, the results were striking. The Paper2Agent tool set could run analyses from the published work, answer questions about the findings, and handle new data. It did all of this more accurately than the same AI agent working from the original paper and code, and it did so more than twice as fast. Anyone who has tried to rerun another researcher’s code will understand why. The typical experience involves downloading a tangled bundle of files, guessing at dependencies, confronting version conflicts, and eventually sending an email to the authors and hoping they remember what they did. Paper2Agent short-circuits much of that frustration. It packages the code into reusable pieces, checks that those pieces work against the paper’s own reference outputs, and then supplies step-by-step instructions for an AI assistant to follow. The agent can execute the workflow without having to reinvent it each time. That is not a minor convenience. It lowers the barrier to building on other people’s work, which is one of the central ways science advances. Instead of spending precious hours on thankless grunt work, researchers could focus on more interesting questions. The speed and accuracy improvements suggest that this kind of active paper is not only possible but practical. At the same time, the 26 failures are a useful warning. They show how much work remains to make research assets genuinely shareable, and they put pressure on journals and funders to require better data and code practices.
Peter Clark, a senior research director at the Allen Institute for Artificial Intelligence in Seattle who was not involved in developing Paper2Agent, sees both the appeal and the significance of the idea. Researchers already share code on platforms like GitHub, but sharing code is not the same as making it usable. Clark says he often has to “grub around” for half a day just to figure out how to get something running. It is a familiar experience: the repository has no README, the environment is outdated, or the key script is buried somewhere unexpected. General-purpose AI assistants, like Claude, can help, but they tend to generate new code on the fly, improvising solutions that may or may not match the original approach. Paper2Agent takes a different path. It packages code associated with a paper into reusable tools, tests those tools against reference outputs, and then writes step-by-step instructions telling an AI assistant which tools to use and in what order. This is more disciplined. Instead of a smart assistant making things up as it goes along, it follows a validated path. For Clark, that streamlines and standardizes the process of using AI to interact with scientific research. He describes Paper2Agent as “an exciting step” toward papers that actively help scientists. The phrase “active paper” captures a genuine shift in thinking. A traditional paper is a report; a Paper2Agent-style paper is more like a well-documented, tested software library wrapped in a scientific narrative. It can answer practical questions like “What happens if I apply this method to my dataset?” and “Which variables matter most?” without requiring the user to parse dozens of dense equations or trace through the original code. That could be especially valuable in fields where computational methods are complex and change quickly. It also helps with reproducibility, because the tool set has been checked against the paper’s own expected outcomes before it ever reaches the user. The system is still young, but it points toward a future where the paper is part of the research process, not just a record of it.
To James Zou, this transition is a natural next step in how human beings share knowledge. For most of history, knowledge was tied to physical objects: stone tablets, papyrus, paper, and eventually digital pages. Each new medium made information easier to copy, store, and distribute. But in every case, the knowledge itself was passive. It waited for a human reader to find it and make use of it. Zou wants to move to active formats, where knowledge can do work on its own. A paper-agent can help one scientist quickly use another person’s code and methods. It can also search for other paper-agents on similar topics, making connections that might not be obvious from reading abstracts or titles. If two researchers are using overlapping techniques, the agents could in principle alert their human authors about a possible collaboration. That is an intriguing idea: knowledge that not only responds to questions but also reaches out to other knowledge. Zou says his team has already created “over 100 different paper-agents” from their own research. And he says these agents have already helped him find other scientists to work with. It is a small glimpse of what an active scientific literature might look like. Instead of hoping that a relevant paper crosses your desk, the papers themselves could find each other, and by extension find you. That does not mean the human element disappears. It means the drudge work of staying current becomes easier, and the creative work of connecting ideas becomes more possible. There is something almost poetic about the thought: a paper you wrote years ago could one day strike up a conversation with a paper across the world, and neither of you would know until the two human authors get an unexpected email. That is the kind of future Zou is trying to build. It may not happen overnight, but the pieces are already in place.
Not everyone is ready to embrace this vision without qualms. Clark points out that Paper2Agent specializes in helping scientists reuse methods and code. It is not yet the kind of general-purpose virtual author that can answer open-ended questions about a paper’s motivations, limitations, or broader meaning. The range of what it does well is real but narrow. He also raises a more practical worry: what happens when millions of low-quality papers are transformed into agents capable of pestering scientists? The internet already has enough noise; adding a swarm of automated paper-salespeople could make things worse rather than better. Clark remembers Clippy, the much-maligned Microsoft Office assistant, and he does not want a scientific version of that popping up every time you open a document, offering unsolicited help. The fear is not about the technology itself but about how it is deployed. If every half-baked preprint gets a chatbot that sends “have you considered citing me?” messages, the entire idea could become annoying and toxic. But Clark also admits he would love a well-designed, “non-annoying agent” that could pop up and tell him about work he should be paying attention to. The line between helpful and irritating is thin. It will depend on quality control, user control, and the norms that the scientific community develops. If agents are built carefully, tested thoroughly, and allowed to operate at the user’s discretion, they could feel like trusted research assistants. If they are built recklessly, they will feel like Clippy: a well-intentioned nuisance that everyone tries to disable. The technology is not inherently good or bad; it is a tool, and the outcome will depend on choices that researchers, publishers, and developers make now. Used wisely, it could open doors; used carelessly, it could become another source of distraction in an already crowded world.
Even with all the promise, Clark says there are things no paper-agent can replace. For more open-ended questions, he would still want to talk to the people who did the research. “The reason I would write to a corresponding author is for precisely the things they didn’t put in the paper.” That is a lovely and important point. A paper is a cleaned-up version of the messy process of science. It contains the findings the authors chose to present, but not all the dead ends, the hunches, the failed experiments, or the private doubts. Those live in the minds of the researchers. A paper-agent can tell you how to run the code, but it cannot tell you why the team chose one approach over another, what they would do differently next time, or what they sense but cannot yet prove. That kind of context requires a conversation with a human being. The best future may be one in which Paper2Agent handles the mechanics while researchers handle the meaning. If an agent can run a new analysis on your data and summarize the results, it saves time. But the real scientific insight often comes from asking why, not just how. The paper-agent can make that conversation happen sooner and better, by freeing researchers from the grunt work that currently slows them down. It may also create new collaborations, by pointing out whose work is worth a real conversation. The dream is not to turn science into an automated assembly line. It is to make scientific knowledge more useful, more discoverable, and more connected, while leaving the curiosity, judgment, and personal connection to the people who actually do science. If Paper2Agent moves the field in that direction, it will be a meaningful step forward. But the last word should probably belong to the corresponding author, because there will always be things that never make it into the paper.
