18 SEP 2026 — A Stanford team pointed an automated framework at 100 computational-biology papers and got 74 working AI agents out of it. The 26 that failed are the more useful result.
Paper2Agent was published in Nature on 16 September. It reads a paper and its codebase, builds a server other AI agents can call, and turns a published method into something a researcher can run by asking for it in a sentence.
What the framework does
The output is a Model Context Protocol server. MCP is the interface standard that lets a model call tools and query data through one common contract instead of a bespoke integration per system, and it is becoming the default way to connect a model to things it does not own.
Paper2Agent analyses the manuscript and the repository using multiple agents, constructs the server, then generates and runs tests against it to harden what it built. The finished agent exposes the paper's tools, datasets and workflows as callable functions, so an AI coding agent can reproduce the original analyses and apply the method to new data.
Three conversions carry the demonstration: AlphaGenome for genomic variant interpretation, ScanPy for single-cell transcriptomics and TISSUE for spatial transcriptomics. We covered AlphaGenome's atlas of nine billion variant predictions earlier this month, and the distinction that mattered there applies here too: an agent wrapping a predictor serves predictions, not measurements.
Seventy-four out of a hundred
The evaluation spans 136 papers across three groups, including the 100 computational-biology papers. Seventy-four of those became agents.
The conversion rate is the number to carry, not any accuracy score. A 74 per cent success rate on real published work says the approach works outside a demonstration. It also says a quarter of the field's output is not runnable even by a well-resourced team that built the automation.
Why the others failed
The reasons for failure are mundane, which is the point: incomplete codebases, missing documentation, and environment configurations that could not be resolved.
None of those is a limitation of the model doing the reading. They are properties of how research code is published: a repository that omits the script that produced figure three, a dependency list that no longer installs, a pipeline documented in a graduate student's memory. Paper2Agent did not fail to understand those papers; it failed to run them, exactly as any other researcher would.
The 26 failures, then, measure the reproducibility of the research rather than the capability of the tool. An automated framework attempting every paper in a corpus produces something the field rarely gets: a consistent, unsentimental count of how much published work is actually executable by someone who was not in the lab.
What the authors will not claim
The authors set the ceiling themselves. They describe Paper2Agent as "a tool for augmenting scientific discovery and improving access, reproducibility and reuse of papers, rather than as an autonomous or authoritative source of scientific conclusions."
That sentence is doing specific work. An agent that speaks in the voice of a paper invites the reader to treat its answers as the paper's answers, and a language model wrapped around a codebase can still produce a confident sentence the underlying method does not support. The tests Paper2Agent generates check that the tools run; they do not check that the conclusions drawn from them are sound.
James Zou has been similarly direct about data. "If the user has sensitive data (e.g. protected health information) that they don't want to send to an LLM then they should exclude that data from P2A," he said — which is a plain statement that the convenience of asking a paper a question routes your data through a model provider.
The year in between
One detail matters for anyone reading this as a launch. The preprint went up in September 2025 and was revised the following month, so the journal version arrives twelve months on. This is not a new capability appearing overnight but the peer-reviewed form of something that has been available and improving for a year, which is a better reason to take the conversion figures seriously than novelty would be.
What to watch
The durable question is whether the failures get fixed. If the 26 per cent that resisted conversion shrinks because journals begin requiring runnable environments rather than because the tool gets cleverer, that is the durable outcome, and it benefits every human reader too.
Zou has said he hopes to build a platform where paper agents talk to one another. That is the ambitious version of this idea and the one with the least evidence behind it. Agents exchanging outputs compound each other's errors as readily as their findings.
Cost is the other open question. A paper agent answers by calling a commercial model, so the cost sits with the asker and the dependency with whichever provider it was built against. An access tool that attaches a per-question cost has moved a barrier rather than removed it.
View author profile → · Editorial policy
