09-16-2026, 06:40 PM
The Nature paper introduces Paper2Agent, a framework that converts a conventional scientific paper—along with its code, datasets, supplementary material and workflows—into an interactive AI agent. Instead of treating a paper as a static document that humans must read and then implement themselves, Paper2Agent creates a “virtual corresponding author” that can answer questions, execute the paper's methods, analyse new datasets and reproduce published results through natural-language instructions.
It uses the Model Context Protocol (MCP) to package executable tools, resources and workflow instructions, while an automated testing process checks that the generated tools reproduce the original paper's results. In one example, an AlphaGenome agent generated 22 validated tools in about 45 minutes and achieved very high accuracy on both tutorial-based and previously unseen genomics tasks.
The researchers tested the approach at larger scale on 100 computational-biology papers, successfully agentifying 74 of them and producing 593 validated tools. Across benchmarks, Paper2Agent generally outperformed simply giving an AI coding agent direct access to the paper and repository, while also reducing cost and execution time. More importantly, several paper agents could collaborate with one another: in a psoriasis genetics case study, agents based on AlphaGenome and two experimental studies combined their evidence to prioritize a candidate causal gene and propose a new analysis strategy.
The authors nevertheless emphasize that Paper2Agent does not eliminate researchers: humans still need to choose research directions, evaluate evidence and validate hypotheses experimentally. They also identify limitations involving incomplete code, missing data, software dependencies, security, intellectual property and maintaining agents as underlying research software changes. The broader idea is a major shift from “papers that describe research” to executable, interactive research objects that can actively participate in scientific workflows.
ARTICLE
It uses the Model Context Protocol (MCP) to package executable tools, resources and workflow instructions, while an automated testing process checks that the generated tools reproduce the original paper's results. In one example, an AlphaGenome agent generated 22 validated tools in about 45 minutes and achieved very high accuracy on both tutorial-based and previously unseen genomics tasks.
The researchers tested the approach at larger scale on 100 computational-biology papers, successfully agentifying 74 of them and producing 593 validated tools. Across benchmarks, Paper2Agent generally outperformed simply giving an AI coding agent direct access to the paper and repository, while also reducing cost and execution time. More importantly, several paper agents could collaborate with one another: in a psoriasis genetics case study, agents based on AlphaGenome and two experimental studies combined their evidence to prioritize a candidate causal gene and propose a new analysis strategy.
The authors nevertheless emphasize that Paper2Agent does not eliminate researchers: humans still need to choose research directions, evaluate evidence and validate hypotheses experimentally. They also identify limitations involving incomplete code, missing data, software dependencies, security, intellectual property and maintaining agents as underlying research software changes. The broader idea is a major shift from “papers that describe research” to executable, interactive research objects that can actively participate in scientific workflows.
ARTICLE
┌────────────────────────────────┐
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘