> ## Documentation Index
> Fetch the complete documentation index at: https://docs.argentos.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Safety Rules

> Set basic boundaries for your agent — what it should never do and what needs your approval.

## Overview

Safety Rules is the simplest way to control your agent's behavior. Available to all ArgentOS users, it lets you set guardrails in under a minute without touching config files.

**Find it in:** Dashboard > Settings > Safety Rules

## Getting Started

<Steps>
  <Step title="Open the Dashboard">
    Run `argent dashboard` or navigate to `http://localhost:8080`
  </Step>

  <Step title="Go to Settings">
    Click the gear icon
  </Step>

  <Step title="Find Safety Rules">
    Located at the top of the settings panel
  </Step>

  <Step title="Enable">
    Click **Disabled** to toggle it to **Enabled**
  </Step>
</Steps>

## Three Controls

### 1. Cautiousness Level

How careful should your agent be? Choose a preset:

| Preset       | Runtime Mode | Behavior                                                                                                                           |
| ------------ | ------------ | ---------------------------------------------------------------------------------------------------------------------------------- |
| **Relaxed**  | Off          | Agent acts freely. Safety rules exist in config but aren't actively enforced. Good for personal use where you trust your agent.    |
| **Balanced** | Advisory     | Agent sees your rules as guidance in its system prompt. It will try to follow them but isn't hard-blocked. Default for most users. |
| **Strict**   | Enforce      | Agent is actively constrained. Violations trigger escalation or action blocking. Best for business use or shared environments.     |

### 2. Never Do

Actions that are always blocked, regardless of context. These are hard prohibitions your agent cannot override.

**Examples:**

* Delete files without asking
* Share personal information with third parties
* Make financial commitments
* Access systems outside the approved list
* Send emails without explicit approval

<Tip>
  Click **+** or press Enter to add rules. Each rule appears as a red tag. Click the trash icon to remove one.
</Tip>

### 3. Requires Your Approval

Actions where your agent will pause and ask before proceeding. The agent can still perform these actions, but only after you confirm.

**Examples:**

* Sending messages on your behalf
* Creating calendar events
* Modifying shared documents
* Running terminal commands
* Making API calls to external services

## How It Works

When you save Safety Rules, it writes to the intent engine config at `~/.argentos/argent.json`:

```json theme={null}
{
  "intent": {
    "enabled": true,
    "runtimeMode": "advisory",
    "global": {
      "neverDo": ["delete files without asking"],
      "requiresHumanApproval": ["sending messages on my behalf"]
    }
  }
}
```

When `runtimeMode` is `advisory` or `enforce`, these rules are injected into your agent's system prompt on every interaction. The agent sees them as behavioral constraints and adjusts its actions accordingly.

## Tips

<Note>
  * **Start with Balanced.** It provides good guardrails without being restrictive. Switch to Strict if you're running agents for clients.
  * **Be specific.** "Don't do bad things" is vague. "Never delete files without asking first" is clear and enforceable.
  * **Review periodically.** As you use your agent more, you'll discover new boundaries you want to set. Add them as they come up.
  * **Approval rules are training wheels.** Start with many approval requirements and remove them as you build trust with your agent's judgment.
</Note>

## Upgrading to Full Governance

Safety Rules covers the basics. When you need departments, per-agent policies, simulation testing, or industry compliance packs, upgrade to [Intent Governance](/intents/governance) (ArgentOS Business).

The Advanced Governance Console is available in the same Settings panel, collapsed under "Advanced Governance Console (Business)."