Jiacheng Miao and James Zou want journals to publish executable agents beside papers, code and data, with researchers still judging the conclusions.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Primary source: [X](https://x.com/james_y_zou/status/2100257219384332333)
Why it matters #
Paper2Agent proposes a concrete publishing standard for AI-era science: methods that can be queried, tested and reused. Its 26% conversion failure rate also exposes brittle research code.
Jiacheng Miao (@Jiacheng_Miao) and James Zou (@james_y_zou) published Paper2Agent in Nature on September 16th, giving researchers an automated way to turn a scientific paper, its code and its supporting data into an AI agent that can explain the work, rerun analyses and apply its methods to new questions. The peer-reviewed publication is new, though the underlying preprint first appeared on September 8th, 2025. Joe R. Davis, Yaohui Zhang and Jonathan K. Pritchard co-authored the work.
Miao designed and developed the framework after joining Stanford as a postdoctoral researcher in June 2025. His route into agentic science ran through statistics and biomedical data science: he earned a statistics degree at Nanjing University, followed by a master's in statistics and a doctorate in biomedical data science at the University of Wisconsin-Madison. At Stanford, he works with Pritchard and Zou on AI co-scientists and the genetics of complex diseases. Zou, a Stanford associate professor of biomedical data science with appointments in computer science and electrical engineering, co-conceived Paper2Agent with Miao and supervised the work.
Their starting point is a familiar failure in computational research: a published result can be readable while remaining painful to use. A researcher still has to locate the right repository, install fragile dependencies, reconstruct undocumented assumptions and work out which inputs produce which outputs. "Papers are written for humans," Zou wrote in the announcement thread. The researchers wanted the agent to run the work inside a virtual lab, gaining experience with the methods instead of treating the PDF as another document to summarize.
The paper becomes an MCP server
Paper2Agent uses a six-stage, multi-agent pipeline. It locates the paper's repository, builds an isolated software environment, identifies tutorials, executes them, extracts reusable functions and tests those functions against the original outputs. The validated tools are then assembled into a Model Context Protocol, or MCP, server that a coding agent can call through natural-language instructions.
That structure is the central technical choice. A language model reading a PDF can answer questions about the text, but it does not automatically inherit the paper's working environment, datasets, executable examples or analysis steps. Paper2Agent packages those materials as tools and structured resources. Failed functions are excluded from the final MCP server after repeated test attempts, leaving a record of what ran and what did not.
The open-source repository now provides a skill that coding agents including Claude Code, Codex and the Gemini command-line interface can use to convert a paper and its associated repository. The MIT-licensed code had about 2,500 GitHub stars as of September 17th, a measure of developer interest rather than evidence of scientific adoption.
Miao and his collaborators demonstrated Paper2Agent with agents built around AlphaGenome, Scanpy and TISSUE, three computational tools used in genomics and single-cell analysis. The agents reproduced published analyses and handled new queries. In one supplementary case, an AlphaGenome agent worked with an ADHD genome-wide association dataset agent to prioritize a variant that may affect the splicing and expression of MPHOSPH9. The paper treats that result as a candidate mechanism requiring experimental validation.
A separate demonstration combined three paper agents to analyze a psoriasis-associated genetic locus. The agents connected predictions from AlphaGenome with two independent perturbation datasets and prioritized GPR137 as the probable causal gene. The analytical procedure crossed methods and datasets that had originally been published separately, showing why Miao and Zou see paper agents as collaborators rather than chat interfaces attached to individual PDFs.
A reproducibility test hiding inside the agent
Paper2Agent's large-scale evaluation also exposed the condition of research software. The researchers ran the pipeline on 100 computational biology papers sampled without filtering for documentation or repository quality. Paper2Agent successfully converted 74 into validated agents. The other 26 failed because of problems including incomplete code, missing documentation and software environments that could not be resolved.
That 26% failure rate is one of the paper's most useful findings. Agent creation becomes a reproducibility audit: software that cannot be installed, tested and generalized is difficult to package into a reliable agent. Paper2Agent can repair some broken dependencies, file paths, typographical errors and outdated API calls, but automation cannot reconstruct code or data that was never published.
The authors also draw a firm boundary around the benchmark results. Accuracy measures whether Paper2Agent faithfully executes a paper's workflows and matches expected outputs. It does not establish that the underlying scientific conclusion is correct. Several interpretations can be defensible in open-ended analysis, and researchers remain responsible for selecting questions, reviewing evidence and deciding which conclusions survive scrutiny.
From code availability to agent availability
Miao and Zou's longer-term proposal is a new artifact in the publishing process. Journals already ask authors to provide data-availability and code-availability statements. The Nature paper argues that an "agent availability" section could tell readers whether a study's methods have been packaged into an interactive, executable agent.
That proposal would shift work onto authors and maintainers. Paper agents need updates when dependencies, APIs and upstream repositories change. Executable research also creates security, intellectual-property and attribution questions that a static PDF largely avoids. An agent with shell access and scientific tools carries a different risk profile from a downloadable appendix.
Paper2Agent is currently an academic, open-source research effort rather than a commercial software company. That gives Miao and his collaborators room to test a publishing format before deciding whether it belongs in journal infrastructure, research labs or a dedicated product. Miao's broader research agenda already pairs enthusiasm for AI co-scientists with work on how agentic systems can distort scientific conclusions without guardrails. Paper2Agent puts the constructive half of that agenda into working code: let researchers interrogate and reuse published methods, while keeping scientific judgment with the humans who are accountable for the result.