Instructions
AGENTS.md files hold instructions that e adds to the system prompt. e wraps
each one as project instructions, with the file's path.
On this page
Where e looks
| File | Scope | When it loads |
|---|---|---|
~/.e/AGENTS.md | Yours, for every project. | Always. |
<workspace>/AGENTS.md | The project's. | Once the directory is trusted. |
<workspace>/<dir>/…/AGENTS.md | Nested, for one directory. | The first time a tool touches a path under that directory. |
Trust a workspace with /trust, or with e trust [dir] when there is no
terminal. A run refuses an untrusted workspace outright, because an untrusted
repository could otherwise steer the agent.
Files are capped at 32 KiB each.
Nested instructions
Nested files keep a monorepo's rules local. services/api/AGENTS.md is not in
context until the agent touches services/api/. Then it loads, without anyone
asking.
- A nested file loads the first time a tool reads, writes, edits, or searches a path under its directory.
- It arrives as a message in the conversation, not in the system prompt.
- The nearest file arrives last, so it reads as the most specific.
- Each nested file loads once per session.
- Nested files load only in a trusted workspace, and only for paths inside it.
Trust is the precondition
e runs only in a directory whose own instructions you have accepted. A session that started untrusted would work in a repository with no say in what the model was told.
- On your first visit, e asks with the terminal's trust panel.
- Declining exits and records nothing, so the next launch asks again.
e trust [dir]records the answer for a session with no terminal.e untrust [dir]refuses that directory deliberately.
Trust extends to everything inside a trusted ancestor. For that reason e never records your home directory as trusted.
Related guides
Reasoning, skills, and prompt templates have their own guides:
e docs skills and e docs prompt-templates.