agents

name	agents
description	Guide for working with AGENTS.md — the open standard for AI coding agent configuration. Use when creating, updating, or optimizing AGENTS.md files, or when onboarding a project to multi-agent workflows.

Agents.md — Universal AI Agent Configuration

AGENTS.md is an open standard for guiding AI coding agents. It works across 25+ agents including Claude Code, GitHub Copilot, Cursor, Devin, Windsurf, Cline, Aider, and more.

What AGENTS.md Does

• Provides a single, predictable location for project-specific AI instructions
• Works universally — any AI agent that reads markdown can use it
• Supports monorepo patterns with nested files (closest file wins)
• Complements platform-specific files (CLAUDE.md, .cursorrules, GEMINI.md)

When to Use AGENTS.md

Use AGENTS.md as the universal baseline that works with any agent. Use platform-specific files for features unique to that platform:

File	Scope	Best For
AGENTS.md	Universal — all AI agents	Project overview, build commands, code style, architecture, testing
CLAUDE.md	Claude Code only	Claude-specific features, memory-bank, skill references
.cursorrules	Cursor only	Cursor-specific behaviors
GEMINI.md	Gemini CLI only	Gemini-specific context

Rule of thumb: If an instruction applies to any AI agent working on your project, put it in AGENTS.md. If it's platform-specific, put it in the platform file.

File Structure

AGENTS.md is standard Markdown. No required schema — use headings that make sense for your project. Recommended sections:

# Project Name

## Project Overview
[What this project does, its purpose]

## Build and Test Commands
[How to install, run, test, build]

## Tech Stack
[Languages, frameworks, runtime, package manager]

## Code Style
[Conventions: TypeScript, semicolons, quotes, formatting]

## Architecture
[Patterns, directory structure, key design decisions]

## Testing
[Test framework, coverage requirements, how to run tests]

## Dependencies
[Key packages and their purposes]

## Guidelines
[Rules for AI agents: what to do, what to avoid]

Monorepo Support

In monorepos, place AGENTS.md at each package/app level. The nearest file to the code being edited takes precedence:

monorepo/
├── AGENTS.md              # Root-level defaults
├── packages/
│   ├── api/
│   │   └── AGENTS.md      # API-specific rules (overrides root)
│   ├── web/
│   │   └── AGENTS.md      # Web app rules (overrides root)
│   └── shared/
│       └── AGENTS.md      # Shared lib rules (overrides root)

Generating AGENTS.md with codekit

codekit auto-generates AGENTS.md alongside other platform files:

# Generate for all platforms including AGENTS.md
codekit learn --all

# Generate only AGENTS.md
codekit learn -p agents

# Sync all platform files
codekit sync

The generated AGENTS.md includes:

• Project overview and type detection
• Build/test commands from package manager
• Full tech stack and dependency summary
• Code style conventions detected from the codebase
• Architecture patterns
• Installed skills reference
• Memory-bank session continuity instructions

Best Practices

1. Keep it concise — Agents parse the entire file. Long files waste tokens.
2. Focus on "what" not "how" — Describe conventions and rules, not step-by-step tutorials.
3. Include commands — Build, test, lint, and deploy commands are the most universally useful.
4. Update it — AGENTS.md should evolve with the project. Run codekit sync after major changes.
5. Don't duplicate — If CLAUDE.md or .cursorrules already has platform-specific instructions, don't repeat them in AGENTS.md.

Relationship to Other codekit Files

AGENTS.md          ← Universal baseline (all agents read this)
    ↕ complements
CLAUDE.md          ← Claude-specific (skills, memory-bank instructions)
.cursorrules       ← Cursor-specific (IDE behaviors)
GEMINI.md          ← Gemini-specific (CLI context)
memory-bank/       ← Session continuity (referenced by all files)

All these files are generated and synced by codekit learn and codekit sync. AGENTS.md acts as the common denominator that ensures any AI agent — current or future — can understand your project.