Policy Examples & Best Practices

Once you understand how policies work, the next question is usually: what should mine actually look like? Here are common patterns that work well for most teams.

Read-only by default

For agents that only need to look things up (answer questions, pull reports, summarize data), the safest setup is:

• Allow all read actions
• Deny all write, update, and delete actions

This lets an agent be genuinely useful without any risk of it changing something it shouldn't.

Require approval for anything irreversible

If your agent has legitimate reasons to write or update data, but you're not ready to fully automate it:

• Allow read actions
• Require approval for writes and updates
• Deny deletes entirely, or also require approval

This means a human reviews and confirms any action before it actually happens — useful while you're still building trust in the agent's judgment.

Scoped-down write access

Once you trust an agent for specific, well-understood tasks, you can narrow policies to exact actions rather than broad categories:

• Allow "create a new ticket in Onboarding Pipeline"
• Require approval for everything else

This gives the agent room to do its specific job without opening the door to anything unplanned.

Combining conditions

Policies can be layered. For example, you might allow writes generally, but add a more specific rule that denies deletes in a particular pipeline or blocks actions outside business hours. The most specific matching rule always wins.

A good habit

Start restrictive and loosen over time, rather than starting permissive and tightening later. It's much easier to grant more trust once you've seen how an agent behaves than to walk back an action that's already happened.