First things first: what is an AI agent?

In about 45 minutes, Paper2Agent built 22 tools for Google DeepMind’s genome model AlphaGenome on a personal laptop, at a cost of US$14. The test shows how the system can turn a research paper into an AI agent that runs its methods.

From passive documents to agents that get things done

Researchers applying a paper’s method often spend days installing the authors’ software and figuring out the data format. Earth.com reported on October 4 that Paper2Agent, developed by Stanford postdoctoral scholar Jiacheng Miao, associate professor James Zou and colleagues, removes that step: users ask in plain language, and the agent runs the paper’s code to answer. The study appeared in Nature.

Paper2Agent's three-agent workflow: set up the environment, package the tools, test them against the paper's results, then bundle everything into an MCP package
Paper2Agent’s three stages

Tools that pass testing are bundled with the paper’s text, data and common analysis steps through the Model Context Protocol (MCP), a shared standard that lets AI access external tools and data. Zou describes MCP as a way to make papers easy for agents to access, “almost like a filing system”.

What the numbers say

In the AlphaGenome test, the system built 22 tools in about 45 minutes without human intervention. It ran on a personal laptop, and the cost was just US$14.

On 30 open-ended research questions requiring several tools to work together, the agent scored 82.7%. Claude, working directly from the code, scored 56.7%; Biomni, another AI research assistant, scored 72.2%.

In the larger test, 74 of 100 computational biology papers were successfully turned into agents. Of the 599 proposed tools, 593 passed testing. Across 300 questions, the agents scored 91.2%, compared with 80.3% when working directly from the code.

Paper2Agent test results: 82.7% success rate, 74 of 100 papers converted, 593 of 599 tools passed, 22 tools built in about 45 minutes, US$14 cost
Key numbers

Three agents team up to find a psoriasis gene

Another test assigned work to three paper agents. The AlphaGenome agent predicted that, in CD4+ T cells (white blood cells that regulate immune responses), psoriasis-linked loci most affect GPR137. Cross-checking with agents built from two other papers left GPR137 as the only match among five candidate genes, and only in stimulated cells. The analysis used existing data and involved no new experiments.

The limits are clear too

The authors stress that agents cannot take the place of independent scientists: researchers still choose the direction and evaluate the evidence. Anyone publishing an agent must also maintain its underlying software. Security, ownership and academic credit remain unresolved. The authors suggest journals may eventually ask papers to state whether an agent is available.

What it means for Taiwan

Paper2Agent could not convert 26 papers because code or data were missing, or their environments could not be installed. These failures offer a lesson for research institutions. Taiwan’s research institutes and universities produce many project reports and experimental datasets, yet code and data often stay on personal computers and lose support when projects end. For agents to call research outputs as they did the 22 AlphaGenome tools, knowledge management needs to preserve code, data and runtime environments together.

On the left, an old file room piled high with PDFs and cardboard boxes; on the right, a lab where code, data and environments are neatly in place and AI agents can work with them directly
Delivering research outputs
  • Deliver code, data and runtime environments together with reports as research outputs.
  • Set the rules on access to internal data, maintenance responsibilities and credit attribution first.
Sources (3)

Sources