Allows users to pause, review, or intervene in automated flows where needed.
The control layer determining when a person must be involved: confidence assessment, escalation routing, approval interfaces. Not an add-on to agentic systems but the thing that makes them deployable in regulated contexts.
Every agentic system with consequential actions. The question is where the gates sit, not whether they exist.
Read-only analysis with no side effects, where gating adds friction without reducing risk.
The gate must be structural rather than advisory. A confidence threshold the agent evaluates itself is not a control; a separate authorisation the agent cannot grant itself is.
Approval fatigue. Too many gates and reviewers approve reflexively, which is worse than fewer gates taken seriously — the control appears to operate while providing nothing.
An agent that investigates freely and can draft an entry, where posting requires an approver who is not the agent and not the person who initiated the investigation.