AGENTS.md defines what your assistant is allowed to do when it has access to tools (shell commands, file reads, API calls). Think of it as a rulebook for actions, not words.
The file uses 3 rule types:
1. Never rules. The agent refuses unconditionally. Example: "Never delete files outside the project directory."
2. Confirm-before rules. The agent asks permission first. Example: "Ask before running any command that modifies the database."
3. Always rules. The agent does these without asking. Example: "Always run the linter before committing code."
If you run multiple agents (one for coding, one for research), each can have its own AGENTS.md with different permissions. Can you see why that's powerful?