.agentic.md

A lightweight, opt-in convention for open source projects to express how AI agents may (or may not) contribute — inspired by robots.txt, built for sustainable collaboration.

We think that every open-source project needs a clear agent policy, and these policies should be quantifiable.

You can refer .agentic.md in AGENTS.md or CONTRIBUTING.md.

You can also add a PR template and a Github action to check these rules.

Distinction Between AGENTS.md and .agentic.md

• AGENTS.md: Designed specifically for AI agents.
• .agentic.md: Designed specifically for human developers.

A suggested document

.agentic.md

# For all contributors
## target
  contributor / reviewer / maintainer
## action
  issue / pr / review / push / release
## Allowed
  L1 - L2, Using an AI assistant while programming is permitted.
## Cautiously Allowed
  L3 - L5, AI-generated code may be accepted only if it has been thoroughly reviewed by a human before publication. Such submissions will be reviewed at the team’s discretion, subject to available capacity.
  In the commit msg, you should add a clear promise, for example:
  Co-authored-by: Claude <noreply@anthropic.com>
  Reviewed-by: lifubang <lifubang@acmcoder.com>
  Signed-off-by: lifubang <lifubang@acmcoder.com>
## Not Allowed
  L3 - L5, Submissions consisting entirely of AI-generated code without prior human review are strictly prohibited.

# Definition of AI Usage Levels
  ## L1
  AI Assistant: The developer uses an AI tool (e.g., chat-based assistants like Qwen, ChatGPT) to answer questions, explain concepts, suggest debugging strategies, or provide small code snippets during interactive development. The human remains fully in control of design, logic, and implementation.
  ## L2
  AI Copilot: The developer uses real-time AI pair programming tools (e.g., GitHub Copilot, CodeWhisperer) that suggest lines or blocks of code as they type. These suggestions are actively reviewed, edited, and accepted by the developer in real time within their IDE.
  ## L3
  Vibe Coding: The developer provides vague, high-level prompts (e.g., “make a login page” or “add user auth”) and accepts large chunks of AI-generated code with minimal scrutiny, relying on “vibes” rather than deep understanding. The output may be integrated with only superficial edits.
  ## L4
  Spec Coding: The developer writes a detailed technical specification or pseudocode and asks an AI to generate a full implementation from it. While the spec is human-authored, the actual code is primarily AI-produced.
  ## L5
  Agentic Coding: An AI agent (or chain of agents) autonomously designs, implements, tests, and potentially submits code with little to no human intervention beyond an initial goal statement (e.g., “build a REST API for todo app”). The human does not meaningfully engage in the coding process.