Stanford's Paper2Agent, published in Nature, turns a paper into a chat agent

Why this matters

Most papers are still PDFs. Paper2Agent only helps when the authors already published the code and the data. Where that exists, the pitch is that you can ask the paper to rerun a figure or apply the method to new data instead of re-implementing it yourself.

The failure mode is ordinary and useful to know: incomplete code, missing docs, and environments that will not build. A paper without a working repository does not become an agent.

What we know

• ✓ Nature published the Paper2Agent paper, DOI 10.1038/s41586-026-11044-y, from a Stanford team led by Jiacheng Miao and James Zou. The Register reported the Nature publication on 17 September 2026. ^[1][2]
• ✓ The system turns a manuscript plus its code and data into an agent that can discuss the findings, reproduce analyses, apply the method to new data, and collaborate with other paper-agents. ^[2]
• ✓ The Register, citing the authors, reported that 74 of 100 computational-biology papers were successfully turned into agents, with failures blamed mainly on incomplete codebases, missing documentation, or environments that could not be rebuilt. ^[2]

What is unclear

• ? How often an outside lab, not the authors, can rebuild those agents from the Nature paper alone. The 74-of-100 figure is the authors' evaluation, as reported by The Register.
• ? Whether "collaborate with other paper-agents" produces a result a domain expert would trust, or only a plausible conversation. The reporting we used does not include an independent replication.