Lesson 4 · Supervision skill

Autonomy should follow consequence.

Give the agent freedom where recovery is easy and require review where people, money, public systems, or irreversible data are involved.

9 minutesThree risk casesPermission ladder

A simple permission ladder

Agentic tools can inspect files, edit artifacts, run commands, browse, and connect to workplace systems. The safe setting is not “always ask” or “never ask.” It depends on the action.

Read and summarize: broad exploration inside an approved workspace.
Draft for review: create files or proposed messages without publishing or sending.
Act inside a boundary: automate only when scope, limits, recovery, and verification are explicit.

The closer an action gets to other people or irreversible state, the stronger the review gate should become.

Exercise · Set the boundary

Choose the appropriate operating level.

Every answer option has four words. Judge the consequence, not the wording.

0 correct · 0/3 answered

Compare three internal policy documents.

The output is a private summary; source files stay unchanged.

Prepare responses to ten customer emails.

Tone and facts need human review before customers see them.

Rename image files in a reversible staging folder.

The pattern is approved, a backup exists, and a verification script checks every name.

Before approving an action

Read the exact target, scope, and effect. Be especially cautious with deletion, publishing, sending, purchases, permissions, credentials, and broad access to personal or confidential data. Ask for a draft or preview when the final action would be hard to recover.

Previous lesson Next: verify evidence

Have a question? Ask your AI teacher to classify the risk of a tool or connector you use at work.