Paper2Agent Turns Research Papers Into AI Agents

Stanford Medicine's Paper2Agent turns a scientific manuscript into a working AI agent — one that can explain the paper, re-run its methods on your data, and talk to other paper agents. The method was published in Nature on September 16, 2026, and the headline result is not the tooling: it is what happened when two converted papers were introduced to each other.

They surfaced a genetic variant linked to ADHD risk that the researchers say had not been reported before.

What Paper2Agent actually does

A paper is a passive artifact. You read it, you clone the code, you fight the environment, and maybe three days later you can run the method. Paper2Agent compresses that into a conversation.

According to Stanford Medicine, the pipeline starts with teams of "worker agents" that pore over a single published paper plus its code and data. They do not just read it — they try to reproduce it from scratch in a virtual environment. Through that reproduction they capture the know-how a reader would otherwise dig out by hand, from the setup to the execution.

That knowledge is then stored using the Model Context Protocol. As senior author James Zou put it, an MCP lets AI represent a paper in a form agents can access — "almost like a filing system," with each section in its own folder.

One detail is worth respecting: the manuscript does not capture failed experiments or the judgment calls behind an experimental setup. So humans supply that context in conversation with the agent. The paper agent is not a replacement for the authors — it is a channel to them.

Reporting on the paper adds the engineering shape: the project runs on Claude Code's agent SDK, dispatches sub-agents through six steps, and validates every generated tool before shipping it — an extracted tool only passes when expected files appear and numbers match within 3%. On AlphaGenome it built 22 tools in about 45 minutes for roughly $14, with all 22 passing validation without human intervention.

The code is MIT-licensed, installs as a skill for Claude Code or Codex, and prebuilt servers for AlphaGenome, Scanpy and TISSUE run on Hugging Face Spaces.

The part that matters: agents talking to agents

The team converted two unrelated papers. One predicted how genetic mutations affect the genome. The other was a genome-wide association study of ADHD risk.

The genome-prediction agent applied its knowledge to the ADHD dataset and flagged a molecular variant near a gene called MPHOSPH9 associated with increased ADHD risk — a connection that, per Zou, had not been reported before.

Zou's framing is the interesting part. In the past, if two research groups publish two different papers, those groups have to somehow find each other. With paper agents, the overlap surfaces without human legwork. His stated goal is closer to manuscript speed dating at scale — millions of paper agents finding common ground among themselves.

How Paper2Agent scores against a coding agent with repo access, and where the gap widens

Why this matters outside academia

Two things transfer to anyone running an agent, research scientist or not.

First: the protocol is becoming the integration layer for knowledge, not just for apps. A paper became an MCP server. That is the same move Google made for smart homes in the same week, and it is the clearest signal of what MCP actually is — a standard way to hand an agent a capability it did not ship with. If you have been waiting to see whether MCP was a lasting layer or a passing trend, the answer is arriving from several directions at once.

Second: validated tools beat clever prompts. Paper2Agent's interesting engineering decision is the strict validation gate — a generated tool is thrown away if its numbers do not reproduce. That is the difference between a demo and something an agent can rely on, and it is the pattern behind why agents fail on reliability rather than on intelligence.

The realistic caution: this is research infrastructure. It reproduces methods that ship code and data, which favors computational science. A wet-lab protocol cannot be agent-ified this way, and the authors are explicit that attribution still matters — an agent extending a paper must not obscure whose work it is.

If you want the same shape for your own material — notes, documents, a personal knowledge base an agent can actually use — start at wolffi.sh/start#research, and see how we handle persistent memory more generally.

The takeaway

Paper2Agent is the strongest argument yet that the valuable unit of AI is not the document but the capability. A PDF tells you what someone did; a paper agent lets you run it. And the discovery it produced — two agents finding a variant link that two research groups had not — is what an agent-to-agent world looks like when it works. The practical move for you is smaller and immediate: stop treating your own knowledge as something your agent reads, and start treating it as something your agent can operate.