Twenty-two tools in about 45 minutes, for $14. That is what the Paper2Agent framework produced when its builders pointed it at the AlphaGenome paper, and all 22 passed automated validation.
Paper2Agent, a Stanford project, takes a research paper plus its codebase and turns them into an AI agent you can talk to. In plain terms: instead of rebuilding someone’s analysis by hand from a methods section, you get a working toolkit that runs the paper’s methods on your own data.
Think of a recipe that arrives with a cook attached. The cook follows the original instructions, then tries the dish with different ingredients.
How it works
The system extracts the core workflows from a paper, builds them as MCP tools, and tests those tools against the original results. Tools that repeatedly fail validation are thrown out. What survives is connected to an AI agent that takes plain-language requests.
That validation step is the part worth noticing. Computational science has a reproducibility problem, and a tool that checks its own outputs and figures against reference examples goes straight at it.
The AlphaGenome test
The generated AlphaGenome agent scored 98.7% accuracy on 15 queries drawn from the paper’s tutorials and 100% on 15 new ones. On 30 open-ended questions written the way a researcher might ask them, it reached 82.7%.
The authors also linked several paper agents together. AlphaGenome, MPRA-coupled scCRISPRi and Perturb-seq agents were used jointly to investigate a psoriasis-associated variant, and they prioritized GPR137 as a probable causal gene. The work is credited to Jiacheng Miao and colleagues, including Jonathan Pritchard, Joe Davis and Yaohui Zhang.
Where it stops
One commentator who flagged the project was careful about limits. It looks most convincing in narrow computational fields, where papers already spell out enough procedure to become a workflow. That, the commentator noted, is not the same as saying every paper turns into a trustworthy agent.
The practical upshot: the 82.7% on open-ended questions is the number to watch as outside groups try it on their own papers.
