# Protect what matters

Agents write to your brains on your behalf. You decide which parts they can only propose changes to, and what instructions they follow while they work.

## Ground truth

Every brain has a `ground-truth/` folder for facts somebody has checked: agreed numbers, current status, decisions that stand. Files there show a **Ground Truth** badge.

An agent can never change ground truth directly. Its change becomes a proposal, and a person on the Approve role decides. People on the Approve role can save there directly, and their save is recorded.

Every brain also protects `routines/`, which holds the instructions for [routines](/assistant/schedule-a-routine). A routine runs with nobody watching, so a person approves every change to what it is told.

## Protect more folders

Each brain lists its protected paths in `.brain/policy.yml`. Each entry is a path pattern with a reason a person can read:

```yaml
version: 1

protected:
  - path: "ground-truth/**"
    reason: "Facts the team has checked. Agents may only propose changes."
  - path: "policies/**"
    reason: "Customer-facing policy. Legal reviews every change."
```

An agent's write to any of these paths becomes a proposal. `ground-truth/` and `routines/` stay protected whatever the file says.

Only admins and the owner can change `.brain/policy.yml`. Ask the assistant to make the change and it is sent to an admin for approval.

## Reaching in from other brains

The same file decides whether an agent working in another brain may reach into this one:

- `allow`: it can read here.
- `ask`: it must ask, and a person approves the request first. Requests appear under **Proposals**, as **Access requests**.
- `deny`: it can't.

```yaml
cross_brain_access: ask
```

Whatever this says, an agent can only reach a brain the person it acts for can reach. To keep a brain from certain people altogether, see [Restrict a brain to certain people](/team/restrict-a-brain).

## The agent contract: AGENTS.md

Each brain has an `AGENTS.md` file at its top level. It tells every agent working in that brain how to behave there: what goes in which folder, how to date a status, what never to write. ako gives it to every agent that works in the brain, including Claude when you [connect it to ako](/connections/connect-claude).

ako writes a starting `AGENTS.md` when you create a brain. You see it as **The agent contract** in the last step of setup. Edit it like any other file.

For instructions that apply across many brains, such as house style, use a [skill](/assistant/add-a-skill) instead.