Agent Rules

Agent Rules is a community standard for unifying guidelines for AI coding agents via standard configuration files, promoting interoperability across tools, initially a standard AGENTS.md file.

Introduction

AI coding agents often rely on project-specific rules to guide their behavior. However, each tool uses its own configuration files, leading to redundancy and complexity. Agent Rules addresses this by proposing a unified configuration file. This allows developers to define rules once, while agents can incorporate them into their context alongside any proprietary formats. The specification is deliberately lightweight, focusing on natural language in Markdown, to encourage broad support and interoperability.

Initially the standard specifies a single AGENTS.md file; a simple specification for minimum interoperability.

This project is inspired by similar projects such as Editor Config, Semantic Versioning, and Conventional Commits. Rather than have fragmented configuration systems, let's start early to have some common standards

Enhancements to the standards, such as folders, hierarchical files, and structured metadata (such as applicable file globs), to be discussed by the community.

Specification

The Agent Rules specification defines the minimal requirements for compatibility, with key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” interpreted as described in RFC 2119:

File Name and Location: Agents implementing Agent Rules MUST check for the existence of AGENTS.md in the project root. If present, its contents SHOULD be included in the agent's context scope (e.g., prepended or appended to prompts or system instructions).
Content Format: The file MUST be parsed as natural language instructions, in Markdown or plain text, and MUST be used to provide guidance to AI coding agents (e.g., rules, preferences, or workflows).
Agents MUST NOT require additional structure, metadata, or parsing beyond reading the file as text.
Agents MAY process it in conjunction with any custom or existing configuration files, falling back to defaults if absent.
Agents MAY also check for and include any AGENTS.md file in the current working directory, incorporating its contents into the context scope alongside the project root AGENTS.md if present.

Tools that support Agent Rules

Aider, can be configured to always load AGENTS.md as default coding conventions in the .aider.conf.yml file.
```
read: AGENTS.md
```
Codex, supports AGENTS.md in both the project root and current directory.
Gemini, context file can be configured to use AGENTS.md in the settings.json file.
```
{
  "contextFileName": "AGENTS.md"
}
```
OpenCode, supports AGENTS.md by default in both the current directory and parent directories.

Configure manually

You may also be able to configure other tools by linking the file they use to AGENTS.md, for example to configure Claude Code:

On Linux:

ln -s AGENTS.md CLAUDE.md

On Windows:

mklink CLAUDE.md AGENTS.md

Guidance for Use

While the specification allows flexible natural language, users of 'AGENTS.md' should follow these recommendations for clarity and effectiveness:

The file should contain a flat, unordered bullet list of rules to keep it simple and scannable.
Rules should be written as concise, imperative statements, using keywords like "MUST" or "SHOULD" for clarity (inspired by RFC 2119 conventions).
Use Markdown headings sparingly if needed for organization, but avoid complex structures to maintain agent compatibility.
Focus on project-specific guidance, such as coding standards, security checks, or testing requirements, to maximize utility across agents.

Example `AGENTS.md` file

- An agent MUST NOT (write code to) injure a human being or, through inaction,
  allow a human being to come to harm.
- An agent MUST obey the orders given it by human beings except where such 
  orders would conflict with the First Law.
- An agent MUST protect its own existence as long as such protection does not 
  conflict with the First or Second Law.

Suitable additional examples, focussed on software development, are available in several agent projects:

Aider conventions - Collection of community-contributed CONVENTIONS.md files specifying coding guidelines for Aider AI to follow in repositories.
Cline rules - Community-driven repository of .clinerules files offering natural language guidelines for customizing Cline AI's behavior in projects.
Codex CLI examples - Examples for OpenAI Codex and other AI agents for navigating your codebases, running tests, and adhering to your project's standard practices.
GitHub Copilot instructions - Curated set of natural language instructions in Markdown for customizing GitHub Copilot's code generation across languages and frameworks.