Principle 09
Human Oversight
People should be able to see what agents are doing on their behalf, approve consequential actions, and intervene at any point.
AX proposalTrust & Control
Why it matters
Delegation does not remove accountability. The person who asked an agent to act is usually the one who lives with the consequences.
Oversight is also how trust is built over time. Visible, reviewable agent activity is what lets people grant more autonomy later.
Examples
Good exampleIllustrative
Before a purchase above the user's threshold, the product returns an approval request that the agent relays; the purchase completes only after the person confirms.
{
"status": "approval_required",
"approval": {
"id": "apr_903",
"reason": "Amount exceeds the 100 EUR limit set for this agent.",
"summary": "Buy 1 × Trail Runner (Black, 42) for 116.10 EUR",
"approveUrl": "https://shop.example/approvals/apr_903",
"expiresAt": "2026-10-06T15:00:00Z"
}
}Bad exampleIllustrative
The product cannot distinguish agent actions from human ones, so the activity log shows the user doing things they never did.
Implementation guidance
- 1Label agent-initiated actions in activity logs and notifications.
- 2Let people configure which actions require their approval.
- 3Make approval requests self-contained so a person can decide without opening the agent's transcript.
- 4Provide an immediate way to pause or revoke an agent.
Evaluation criteria
- Agent actions are labelled and attributable.
- Approval thresholds are user-configurable.
- A person can stop an agent within one step.