For roughly 360 years, scientific publishing has followed the same basic workflow: do the research, write down what happened, publish it, and leave everyone else to figure out how to use it.
Stanford researchers think the last part is due for an upgrade.
A new system called Paper2Agent can take a scientific paper and turn it into an interactive AI agent — not just a chatbot that summarizes the PDF, but something capable of running the paper’s methods, analyzing new data and even collaborating with agents built from other papers.
The work was formally published in Nature, after appearing as a preprint last year. Nature described the resulting agent as a kind of “virtual corresponding author”: a digital version of the paper that can explain what it knows and, in some cases, actually do the work described inside it.
That sounds like an upgraded research assistant.
The more interesting possibility is that the scientific paper itself may be turning into an executable object.
This isn't just ChatGPT for PDFs
We've had AI tools that let you upload a paper and ask questions about it for years. Paper2Agent goes several steps further.
The system ingests the manuscript along with its supplementary material, datasets and associated codebase. A collection of specialized agents then tries to set up the software environment, locate useful tutorials, execute the original workflows, extract reusable functions and test those functions against the results reported by the researchers.
Once that process works, Paper2Agent packages the validated tools into a Model Context Protocol (MCP) server.
MCP is the increasingly common standard that lets AI agents connect to outside tools and data — something we've covered in our beginner's guide to building AI agents. Here, instead of connecting an AI to your calendar or CRM, you're effectively connecting it to the capabilities described by a research paper.
That distinction matters.
A normal AI assistant might read a genetics paper and explain what the researchers did.
A Paper2Agent version of the paper could potentially run the researchers' actual analysis on a new genetic variant.
The Stanford team demonstrated that with AlphaGenome, Google's genomic AI model. Paper2Agent automatically created 22 tools around the system in about 45 minutes, at a reported cost of roughly $14. On the researchers' benchmark, the resulting agent scored nearly 99% on tutorial-derived questions and 100% on novel queries, compared with roughly 83% and 79% for Claude working directly with the original repository.
That points to an important lesson about agents generally: giving an AI a pile of code is not the same as giving it reliable tools.
Paper2Agent does the annoying work in between.
Then the papers started talking to each other
This is where the research gets much more interesting.
The researchers created separate agents representing different papers and allowed them to collaborate on a scientific problem.
In one demonstration, they connected three agents using different genomics methods and datasets to investigate a genetic variant associated with psoriasis. One agent used AlphaGenome predictions, while two others provided separate experimental datasets.
Together, the system produced evidence supporting GPR137 as a probable causal gene associated with the variant. More notably, the agents proposed a cross-dataset analysis strategy that wasn't described in any of the original papers.
That's different from simply retrieving information.
The AI was composing techniques and evidence from several research artifacts into a new analytical workflow.
Humans were still involved: a researcher selected which of the agent's proposed strategies to pursue, and the paper's authors explicitly caution that hypothesis generation and scientific interpretation should remain human-in-the-loop. The system is not being presented as an autonomous authority capable of declaring scientific truth.
But you can see where the researchers want this to go.
James Zou, one of the study's authors, described the goal as being able to “reimagine what knowledge looks like.” Stanford says the group has already generated more than 100 paper agents and eventually imagines much larger networks where agents representing millions of papers could identify useful connections between research that human scientists might never think to combine.
Basically: Tinder for scientific papers, except the matches start running experiments together.
The failed agents might be just as important
There's another finding here that may prove more immediately useful.
Paper2Agent tried to convert 100 computational biology papers into agents.
It successfully agentified 74.
The other 26 frequently failed because the underlying research artifacts were incomplete: missing executable code, unavailable datasets or models, dependency problems, broken environments or scripts that were too specific to generalize.
In other words, the AI ran straight into one of science's oldest problems: reproducing someone else's work is often difficult because the paper alone doesn't contain everything needed to actually do it.
The researchers argue that this could turn “agentifiability” into a practical test of reproducibility.
If an automated system can take your paper, code and data and reliably reconstruct the method, that's a pretty strong indication you've documented the work properly.
If it can't, the failure itself tells you something.
The authors even imagine journals eventually adding an “agent availability” section alongside the increasingly common code and data availability statements. Researchers might publish not just a paper and GitHub repo but a maintained agent capable of demonstrating and applying the work.
That's a much bigger change to scientific publishing than adding another AI search box.
The research paper could become an interface
We're already watching AI change how research is performed.
Projects like Andrej Karpathy's autoresearch have shown how agents can repeatedly modify experiments, run them and use the results to decide what to try next. AI labs are increasingly using agents inside their own research processes.
Paper2Agent attacks the problem from the opposite side.
Instead of asking, How can AI conduct research? it asks:
What if all the research we've already done became usable by AI?
Today, scientific knowledge is largely stored in a format designed for human reading. The PDF is essentially the final product.
Paper2Agent imagines something different: the human-readable paper becomes one representation of the research, while its code, datasets, procedures and reasoning are packaged into something machines can actively use.
There are plenty of reasons to stay cautious. Paper agents need maintenance as software dependencies change. Many papers don't contain enough information to become reliable executable systems. And an agent faithfully reproducing a published method doesn't mean that method — or the conclusions drawn from it — is scientifically correct. The researchers explicitly distinguish reliable execution from analytical validity.
Still, the underlying idea feels important.
The internet made scientific papers searchable.
LLMs made them conversational.
Systems like Paper2Agent are beginning to make them executable.
And once knowledge can act, connect and collaborate instead of merely sitting on a page, the humble research paper starts looking less like the endpoint of scientific discovery — and more like an API for whatever comes next.
