🤖 AI

Agentic Publication Protocol: An Attempt to Modernize Scientific Publication

This paper introduces the Agentic Publication Protocol (APP), a lightweight repository format that modernizes scientific publishing by packaging manuscripts with code, data, and agent-facing instructions to enable AI agents to directly reproduce results, explain methodologies, and support future research.

Original authors: Sirui Lu, Xiao-Liang Qi

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sirui Lu, Xiao-Liang Qi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Static Recipe" vs. The "Living Kitchen"

Imagine you find a recipe for a delicious cake in a cookbook. The recipe tells you the ingredients and the steps. But, it doesn't tell you:

Which specific brand of flour works best.
That the oven in the original baker's kitchen runs 20 degrees hotter than yours.
That the baker tried using chocolate chips first, but they burned, so they switched to vanilla.
How to fix the cake if it sinks in the middle.

In the world of science, a published paper is like that static recipe. It tells you the "what" (the results) but often leaves out the "how" (the messy, practical details). This makes it hard for other scientists to copy the experiment, fix mistakes, or build on the work. They have to guess the missing details, which wastes time and slows down progress.

The Solution: The "Agentic Publication Protocol" (APP)

The authors propose a new way to publish science called the Agentic Publication Protocol (APP).

Instead of just handing out a PDF (the static recipe), APP asks scientists to publish their work as a digital "kitchen" or a "smart box."

Here is how it works, broken down into simple parts:

1. The Package (The Repository)

Instead of just a document, the publication is a complete folder (a Git repository) that contains:

The Paper: The story of what was discovered.
The Code & Data: The actual tools and ingredients used.
The Environment: Instructions on how to set up the "kitchen" (computer software) so it works exactly like the author's.
The "Agent" Instructions (AGENTS.md): This is the most important new part. It is a special instruction manual written specifically for an AI robot (an agent).

2. The "Paper Agent" (Your Personal Research Assistant)

Think of the AI agent as a representative of the author who lives inside the publication folder.

Old Way: You read the paper, get confused about a step, and email the author. They might not reply, or they might forget the details.
APP Way: You talk to the "Paper Agent." It knows the paper, the code, and the hidden tricks the author used.
- You ask: "How did you make Figure 3?"
- The Agent says: "I ran this specific script with these settings. Here is the code. Also, I tried a different setting first, but it failed, so here is what to avoid."

The agent doesn't just summarize the text; it can actually run the code to reproduce the results or guide you through the process.

3. The "Know-How" (Tacit Knowledge)

Scientists often have "tacit knowledge"—things they know but can't easily write down (like "the machine makes a weird noise if you turn it on too fast").
APP encourages authors to teach this knowledge to the AI agent. So, the agent becomes a living memory of the project, holding onto the trial-and-error lessons that usually get lost after a paper is published.

How It Works in Practice

The paper describes a workflow (a set of steps) to turn a normal research project into an APP publication:

Reproduce: Make sure the code actually works and generates the pictures in the paper.
Organize: Put everything (code, data, notes) into a neat, standard folder structure.
Train the Agent: Write the AGENTS.md file. This tells the AI: "Here is the truth, here is the code, and here is how you should answer questions."
Verify: Check that everything is secure and ready for the public.
Release: Publish it as a specific, unchangeable version (like a time capsule) so everyone knows exactly which version they are looking at.

The Results: Does It Work?

The authors tested this system. They took 11 existing scientific papers, turned them into APP publications, and created "Paper Agents" for them.

Then, they asked a simulated student to ask questions about these papers. They compared two helpers:

The General Agent: An AI that just reads the text of the paper and the code files without special instructions.
The APP Agent: The AI trained specifically on the APP protocol.

The Result: The APP Agent was better. It was more accurate, more honest about what it couldn't do, and it pointed to the specific code files that proved the answer. It didn't just guess; it grounded its answers in the actual "kitchen" tools.

The Big Picture

The authors believe that science is moving from a world of static documents to a world of interactive, living research objects.

Before: Research is a network of people passing static papers to each other.
After (with Agents): Research becomes a network where every paper has a "smart assistant" that can talk to other assistants, reproduce results instantly, and help researchers skip the failed paths.

The goal isn't to replace human scientists, but to give them a tool that handles the boring, repetitive, and confusing parts of "re-doing" someone else's work, so humans can focus on the creative, new discoveries.

Summary Analogy

Current Science: You buy a map of a city. It shows the roads, but if a bridge is out, you don't know until you get there.
APP Science: You buy a smart GPS that not only shows the map but also knows which bridges are out, tells you the best detour, and can even drive the car for you if you want. It's the map plus the driver's experience.