Instructions

What it is. The text an agent reads at the start of every session: how to build, test and work in your repository.

Where it goes. AGENTS.md at the repository’s root (~/.agents/AGENTS.md for your own, at user scope).

What each agent gets

Agent	Gets
Codex, Factory Droid, Copilot, Cursor, Devin, any AGENTS.md agent	nothing written: it reads AGENTS.md where it is (shared … in place)
Pi, OpenCode, Kilo Code, Qwen Code	the same, when AGENTS.md is the file they choose; else a copy in the file they read
Claude Code	@AGENTS.md in CLAUDE.md, so it always loads it
Gemini CLI	AGENTS.md added to context.fileName in .gemini/settings.json

Because Claude Code and Gemini CLI are pointed at the file, an edit to AGENTS.md reaches every agent at once, without installing again.

As a tool

A tool’s instructions are text placed between its own markers in each agent’s instructions file, so they sit beside the user’s text without touching it:

<!-- mytool:harness -->
Run `mytool check` before you finish.
<!-- /mytool:harness -->

Command

instructions = "Run `mytool check` before you finish.\n"
# or a file, copied:      instructions = { file = "agent/instructions.md" }
# or a source, read where it is: instructions = { from = "AGENTS.md" }

Rust

Integration::new().instructions("Run `mytool check` before you finish.\n")
// or Integration::new().instructions_from("AGENTS.md")

Node

{ instructions: "Run `mytool check` before you finish.\n" }

Python

{"instructions": "Run `mytool check` before you finish.\n"}

Limits

• Cursor has no instructions file at user scope (its user rules are not files).
• An agent that loads AGENTS.md and CLAUDE.md both (Copilot, Devin) reads text in both; the install warns when it could read something twice.