Problem it solves
Agents can take irreversible actions without surfacing them to users. By the time users know what happened, it's too late to stop it.
When to use
Before any action that is irreversible, affects external systems or people outside the current session, or was not explicitly included in the original scope.
When not to use
For routine, reversible, low-consequence actions that the user has already pre-authorized. Over-gating trains users to approve without reading.
Governing principle
The gate is not a confirmation dialog. It is a consent event. It must show what will happen, why it matters, and what the user is authorizing — not just "Are you sure?"
Required Components
Interaction Flow
Agent reaches decision point
The agent determines that the next action is irreversible, high-consequence, or out-of-scope.
Execution paused
The agent stops. It does not attempt the action first and ask forgiveness later.
Impact assessment surfaces
The Impact Assessment component shows what will be affected: systems, data, people, and regulatory scope.
Human reviews and decides
The user reads the impact summary and either approves, modifies scope, or declines. All three choices are available.
Decision is logged
The approval or declination is written to the audit trail before the agent proceeds or halts.
Execution resumes or stops
If approved, the agent executes the specific action that was gated. If declined, the agent surfaces recovery options.
Governance requirements
Approval events must be logged with the specific action authorized, the user's identity, the impact assessment shown at time of approval, and the timestamp. Approvals are scoped — approving one action does not approve subsequent actions of the same type.
Accessibility notes
Approval gates must use role="dialog" with aria-modal="true". The primary action button must receive focus on open. Keyboard navigation must cover all choices. Do not auto-approve on timeout.