Free practice · ordering · medium

Practice

Order an agent-control workflow

Arrange the controls for a tool-using AI pilot.

Human-control stack

External action requires stronger control than an internal suggestion

  1. GoalBound allowed and prohibited outcomes
  2. AccessGrant least tools and data
  3. TestExercise normal, boundary, and failure cases
  4. ApprovePlace checkpoints before consequential action
  5. RecoverLog, stop, contain, and reverse where possible

Least permission, pre-action checkpoints, limits, logs, stop conditions, and recovery keep a tool-using system inside its approved boundary.

Arrange in the correct order
  1. 1. Log, monitor, stop, and recover
  2. 2. Require approval before consequential actions
  3. 3. Test expected and failure cases in a safe environment
  4. 4. Grant the least required tools and data access
  5. 5. Define the bounded goal and prohibited actions