Advanced level · Module 2: AI Agents and Tool Use · Lesson 3 of 4
Permissions and Human Approval
Limit what agents can do and require human approval for consequential actions.
In this lesson you will
- Apply least privilege to agent tools
- Require approval for sensitive actions
Key concepts
least privilegeread vs write toolsapproval gatesaction logging
The lesson
Give agents only the tools and data they need. Read-only tools are far safer than tools that send messages, change records or spend money.
Put consequential actions behind human approval: the agent drafts the referral, a case worker reviews and sends it. The approval screen should show exactly what will happen.
Scope data access to the user's own cases, and log every action with who approved it.
Step by step
- Classify the capstone tools as read or write.
- Add an approval step for the template filler's output.
- Restrict data access to the user's cases.
- Test that unapproved actions cannot happen.
Practice activity
Implement permissions and approval for the capstone assistant.
Expected result: No write action happens without approval, demonstrated by tests.
Check your understanding
-
What is least privilege for agents?
Show answer
Giving them only the tools and data they need.
-
Which actions need human approval?
Show answer
Consequential ones, such as sending messages or changing records.
Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.