Paper2Agent Turns Research Papers Into AI Agents

A Paper That Can Run Its Own Analysis

For a few hundred years a scientific paper has been a sealed object. You read it, you believe it or you do not, and if you want to actually use the method you clone a repository, fight the dependencies, and hope the tutorial still runs. A Stanford team led by biomedical data science professor James Zou has built a system that collapses that gap. It is called Paper2Agent, it was published in Nature this week, and it converts a paper plus its code into an AI agent you can simply talk to.

• What it builds: a Model Context Protocol server exposing a paper's tools, datasets and workflows to any MCP-capable chat agent
• Scale test: 136 papers processed across three groups, including 100 computational biology papers, of which 74 became working agents
• Case studies: AlphaGenome for genomic variant interpretation, Scanpy for single-cell analysis and TISSUE for spatial transcriptomics
• License: MIT, with the pipeline published on GitHub and prebuilt servers hosted for AlphaGenome, Scanpy and TISSUE

How Does Paper2Agent Actually Work?

The pipeline is deliberately unglamorous, and that is why it works. Paper2Agent runs a chain of specialist agents that locate the paper's codebase, build an isolated environment for it, index whatever tutorials the authors shipped, and then execute those tutorials end to end while recording what came out.

The interesting step is the next one. Rather than trusting the model to describe what the code does, Paper2Agent converts each successfully executed tutorial into a parameterised tool and then checks the tool against the original run. Numeric outputs have to match within three percent. Figures are compared by perceptual hash. Tools that fail are discarded. What survives gets assembled into an MCP server, which is the same open standard that lets AI assistants talk to external tools generally — the plumbing we looked at when OpenAI opened its Agents API.

That validation gate is the whole argument. An LLM asked to summarise a paper will produce fluent text about methods it has never run. An LLM handed a locked, tested tool extracted from the authors' own tutorial is calling the real code. Zou framed the shift plainly, describing papers as static documents for centuries and Paper2Agent as the thing that turns them into agents that can answer questions, apply their methods and collaborate.

What Do the Results Look Like?

The team tested at two scales. The headline case studies are three well-known tools: AlphaGenome for genomic variant interpretation, Scanpy for single-cell analysis and TISSUE for spatial transcriptomics. In each case the resulting agent reproduced the original paper's results and then handled questions the paper never asked. The authors report that the AlphaGenome agent went further and surfaced new splicing variants associated with ADHD risk — not a headline finding on its own, but a demonstration that the agent is running the method rather than describing it.

The breadth test is the more telling one. Per The Register's reporting, Paper2Agent was run against 136 papers across three groups, including 100 computational biology papers, of which 74 were successfully turned into agents.

The failures are as informative as the successes. Of the 100 biology papers, 26 could not be agentified, and the reasons were mundane: incomplete codebases, missing documentation, environments that could not be resolved. Paper2Agent does not repair bad research infrastructure. It rewards good research infrastructure, which is a meaningfully different and rather healthier incentive.

What This Changes for Working Researchers

The honest framing, and one the authors themselves insist on, is augmentation rather than authority. Hallucination has not been solved; results still need checking, and the paper is explicit that an agent should not be treated as an autonomous or authoritative source. Sensitive data should stay out of these pipelines entirely, since queries travel to a language model.

Within those limits the practical effect is large. Methods sections are the slowest-moving part of science — a genuinely useful technique can take years to diffuse because adopting it means a week of setup before you learn whether it helps. A validated agent compresses that week into a conversation. It also gives the reproducibility problem a lever it has not had: a paper whose tools cannot be extracted and re-run is now visibly different from one whose tools can.

This lands alongside a broader move toward AI systems that do scientific work rather than describe it, a direction we traced when a local AI agent matched tumor boards on blood-cancer decisions. More on agent tooling and model releases in our AI coverage.

The code is MIT licensed and public. The most likely near-term outcome is that authors start shipping a paper agent alongside the preprint, because it will be obvious within a year which papers have one and which do not.

Sources: The Register — September 17, 2026; Paper2Agent preprint on arXiv — accessed September 18, 2026; Paper2Agent on GitHub — accessed September 18, 2026.