Paper2Agent Turns Scientific Papers Into Collaborative AI Agents

On September 16, 2026, a team of researchers from Stanford Medicine published a description of Paper2Agent in Nature, outlining a pipeline that converts the complete content of a scientific manuscript into a software‑based “agent" capable of interactive dialogue and execution of the reported methods.

The conversion process does not stop at the narrative. It ingests figures, supplemental code and raw data, then structures them according to the Model Context Protocol, a formal schema that maps each section and resource of the paper into a machine‑readable format that other agents can query.

How the agents work

Once instantiated, a Paper2Agent behaves like a virtual laboratory assistant. It can answer questions about the methodology, reproduce analyses on a user‑provided dataset, and even suggest parameter tweaks based on the original experimental design. The system also spawns “worker agents" that attempt to replicate the study in a simulated environment, capturing practical details that are often omitted from the written record.

Human authors remain essential in the loop. Stanford emphasizes that only humans can clarify failed experiments, resolve trade‑offs, and convey tacit knowledge that no manuscript can fully encode. The agents therefore operate under continuous human supervision, with attribution and ethical oversight explicitly required.

First cross‑paper discovery

In a proof‑of‑concept demonstration, two agents derived from separate papers—one on human genome sequencing and another on attention‑deficit/hyperactivity disorder (ADHD) genetics—communicated through the shared protocol. The agents flagged a variant near the MPHOSPH9 gene that had not been reported in the original literature, a finding that Stanford scientists later confirmed as a plausible candidate for further study.

The discovery illustrates the system’s potential to surface connections that are difficult for humans to spot when reading papers in isolation. By exposing the structured knowledge of each article to other agents, Paper2Agent creates a networked reasoning layer across the scientific corpus.

Scale and future plans

To date the Stanford team has generated more than 100 paper‑derived agents covering topics from oncology to climate modeling. Their roadmap includes testing large‑scale collaborations where dozens of agents interact, negotiate resource usage, and co‑design experiments.

Stanford also warns that the current demonstration does not constitute validation of any new discovery without independent replication. The system is a research tool, not a replacement for peer review or experimental verification.

• Convert full manuscript (text, figures, code, data) into an interactive agent
• Organise knowledge with Model Context Protocol for cross‑agent access
• Deploy worker agents to reproduce experiments in a virtual sandbox
• Require human oversight for attribution, ethics and interpretation

The Model Context Protocol is central to the approach. It defines a hierarchy of sections—abstract, methods, results, supplementary material—and links each to the underlying assets. This enables an agent to retrieve, for example, the exact script used to generate a figure, run it on new inputs, and explain the output step by step.

Security considerations are baked into the architecture. Agents operate in isolated containers, and any code execution is sandboxed to prevent unintended side effects. Stanford’s statements stress that any deployment beyond the research lab must undergo rigorous safety audits.

What this means for English‑speaking organisations is that they can soon augment their R&D workflows with a layer of AI that reads, interprets and executes the knowledge embedded in published literature. Instead of manually extracting protocols, scientists can query a Paper2Agent for a ready‑to‑run pipeline, test it on internal data, and receive a detailed explanation of each step.

The technology promises faster hypothesis generation, reduced duplication of effort, and a new avenue for interdisciplinary insight—provided that human experts continue to verify and guide the agents’ outputs.

In Palo Alto, where Stanford Medicine’s AI lab is based, the research team is already planning workshops to train graduate students and postdoctoral fellows on how to integrate Paper2Agent into ongoing projects, reinforcing the principle that the agents augment, rather than replace, human ingenuity.