Human-in-the-Loop in Production
In production, human-in-the-loop means a real approval queue, notification system, and timeout policy — not just a design principle.
Prerequisites
Overview
Deciding that an action needs human approval is the easy part. Actually implementing it means an approval queue, a way to notify the right person, and a policy for what happens if no one responds in time.
Where It Fits
Agent Requests Action
Approval Queue
Notify Approver
Approve / Reject / Timeout
Key Points
- Approval queue
- Pending actions need to be persisted somewhere durable, not held only in an agent’s in-memory state.
- Notification
- The right human needs to actually be alerted — an approval sitting unseen in a queue is functionally the same as no approval step.
- Timeout policy
- A system needs an explicit decision for what happens if no one responds — auto-reject is usually safer than auto-approve.
Interview Question
What happens if no one responds to an agent’s approval request?
That needs an explicit timeout policy decided in advance — generally auto-rejecting or escalating is safer than defaulting to approval, since silence shouldn’t be treated as consent for an irreversible action. The approval request itself also needs to be durably queued, not just held in the agent’s in-memory state, or it’s lost if the process restarts.
Explain It in 30 Seconds
Implementing human-in-the-loop in production means a durable approval queue, a real notification mechanism so a human actually sees the request, and an explicit timeout policy — usually reject or escalate, not silently approve.