斯坦福把论文做成了 AI Agent，这事上了 Nature 正刊Stanford Turned a Paper Into an AI Agent, and It Landed in Nature

A Stanford team built a tool called Paper2Agent: it automatically packages a paper, together with its code repository, into an interface an AI assistant can call directly — and the work landed in Nature. How it does it, what it costs, and whose work it saves.

Abstract

A Stanford team built a tool called Paper2Agent: it automatically packages a paper, together with its code repository, into an interface an AI assistant can call directly — and the work landed in Nature. How it does it, what it costs, and whose work it saves.

Body

Let's start with a scenario

You're scrolling arXiv and a paper's method looks exactly like what you need. You want to apply its pipeline to your own data.

What follows is roughly two hours of this: clone the repo, build the environment, install dependencies, fight version conflicts, run the tutorial, hit an error, tweak a parameter, hit another error. By the time it finally runs, you've forgotten what you originally wanted to verify.

A group at Stanford has clearly been tormented by the same thing. They built a tool called Paper2Agent, which landed in Nature on September 16, 2026. The paper's title translates roughly to "Reimagining research papers as interactive and reliable AI agents."

In plain terms: it takes a paper plus its code repository and automatically packages the whole thing into an interface an AI assistant can call directly. From then on, you don't read tutorials or run code — you just talk to it in natural language.

The authors gave it a rather nice label: a virtual corresponding author.

So what exactly does it hand you

The input is a paper plus its code repository. The output is an MCP server.

MCP stands for Model Context Protocol, a standard that lets AI assistants from different vendors call external tools in a uniform way. Think of it as a plug specification — build the plug to this spec, and agents like Claude or Codex can all plug into it.

The server exposes three kinds of things.

• MCP tools: the paper's methods wrapped as functions you can call directly
• MCP resources: the manuscript, code links, datasets, figures
• MCP prompts: entire multi-step workflows written in, such as the correct Scanpy preprocessing order

Three are already publicly available, named AlphaGenome, Scanpy, and TISSUE.

Six steps, plus a pass/fail line

The pipeline is built on Claude Code's agent SDK. A central orchestrator directs several specialist sub-agents through six jobs in order.

1. Find and download the code repository
2. An environment manager builds an isolated virtual environment
3. A tutorial scanner indexes the tutorials that actually work
4. A tutorial executor runs the tutorials end to end and records reference outputs
5. A tool extractor converts the tutorials into parameterized MCP tools and hands them to a test validator
6. Tools that pass validation are assembled by the orchestrator into a server

That pass/fail line in the middle deserves its own sentence. A tool has to satisfy two conditions at once to pass: the files that should appear do appear, and its values stay within 3% of the reference output; if figures are involved, it also has to pass a perceptual hash comparison with a Hamming distance under 20. Each function gets at most 6 attempts, and anything that keeps failing is dropped outright and never makes it into the final server.

The whole thing runs on Claude Sonnet 4.

Once installed, using it is a matter of one /mcp entry in Claude Code:

The bill

How much does it cost to run an agent? This set of numbers is refreshingly candid — the authors reported them themselves.

For the AlphaGenome case: 22 tools, about 45 minutes, $14, all 22 passing validation, with no human intervention at any point.

For Scanpy: 7 tools, about 45 minutes, $13, and it matched human researchers on cell counts, gene counts, and top marker genes across 4 public datasets.

The scale test is more aggressive: across 100 bioRxiv computational biology papers, 74 were successfully agentified, and 593 of 599 candidate tools passed validation, with no manual cleanup at any point.

A single query costs $0.20 and takes 1.6 minutes; the comparison condition costs $0.38 and takes 4.3 minutes.

On accuracy, tutorial-derived queries came in at 98.7% (Claude with repository access: 82.7%; Biomni: 37.3%); novel queries at 100.0% (78.7%, 56.0%); open-ended queries at 82.7% (56.7%, 72.2%). The data comes from 5 runs scored by two human experts, with 96.7% inter-rater agreement. When the baseline is swapped for the stronger Claude Opus 4.6, the advantage still holds.

Across 300 questions, the overall score was 91.2%, against 80.3% and 86.3% for the two comparison groups. Ten non-biology papers (including TabPFN, SAM 2, and SAELens) scored 98.1% on 42 execution tasks. Across 26 data-oriented papers, the resource layer scored 89.0% versus 82.0% for the browser approach — while being 34 times cheaper and 15 times faster.

In permutation tests, out-of-scope queries were rejected 100% of the time; when dependency errors, bad file paths, typos, and deprecated APIs were injected, it recovered on its own.

There's something quite grounded in these numbers: 74 out of 100 papers worked. The remaining 26 didn't fail because the tool can't do it — those papers simply didn't leave behind runnable code and tutorials. How broadly it applies is determined by the engineering details other people wrote down.

It really did get used for some research

One case is quite interesting. Researchers chained three agents together: AlphaGenome, an MPRA-coupled scCRISPRi screen (large-scale experiments measuring the effect of regulatory elements), and a Perturb-seq dataset from CD4+ T cells.

AlphaGenome flagged GPR137 at the psoriasis-associated locus rs887314, with an RNA-seq quantile score of 0.997. The AI co-scientist proposed 10 validation strategies, and the researchers picked the signature correlation one. The result: only GPR137 knockdown matched the CRE perturbation signature, and only under stimulated conditions — Spearman 0.613 at Stim8hr and 0.630 at Stim48hr, while the other three candidates (including BAD) showed no significant correlation.

A second thread connected AlphaGenome to a genome-wide association study of ADHD and nominated rs1626703 among 209 candidate loci. That thread remains a hypothesis, waiting on experiments to test it.

One point that has to be spelled out

The paper includes an example involving the LDL cholesterol variant chr1:109274968:G>T. After reanalyzing it, the agent ranked SORT1 as the most likely causal gene, whereas the original AlphaGenome paper emphasized CELSR2 and PSRC1. The GTEx database shows all three genes have significant eQTLs in liver.

What the team itself uses this example to illustrate is this: judging the causal gene at loci like this is inherently hard. It produced a different ranking; it did not correct the original paper.

So whose work does it actually save

By this point the conclusion isn't complicated.

Paper2Agent does a very specific piece of engineering work: it automates away the repetitive, mechanical part of the cost of invoking what's in a paper. It raises the floor of reproduction efficiency, not the ceiling of scientific judgment.

What really determines whether you can put it to use is whether the paper in your hands comes with code and tutorials written clearly enough.

Data sources and notes

The data and conclusions in this article all come from the study and its public materials. The paper went live in Nature on September 16, 2026, under the title Reimagining research papers as interactive and reliable AI agents, DOI 10.1038/s41586-026-11044-y. The preprint was uploaded to arXiv on September 8, 2025, as 2509.06917 — roughly a year between the preprint and the journal publication. The project is released under the MIT license and installs into Claude Code or Codex as a skill. There are five authors: Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard, and James Zou. The three-way comparison data in this article is given in full according to the original reporting, with nothing omitted. This article is a technical summary only, and has no affiliation, authorization, or endorsement relationship with Stanford, Nature, or any institution.

Figure Notes

Cover: reuse the already-generated cover-final.png (official logo overlaid with the Chinese headline).