Human-in-the-loop

Fully autonomous agents are rare in real production. Most production agent systems have humans in the loop at specific decision points, not because the AI isn't smart enough, but because some decisions shouldn't be fully automated even if they could be. Human-in-the-loop isn't a limitation. It's a design choice. Knowing where to place the human is how you ship reliably without becoming a headline.

When the human should pause things

Three interaction models

Pre-action approval: the default for high-stakes

The strongest pattern: agent always proposes, human clicks execute. The agent wrote the email draft; the human presses send. The agent picked the refund; the human approves. This keeps the human as the actor of record, which matters for liability and for user trust. Most users are more comfortable with "AI helped me do this" than "AI did this to me."

Post-action review: for scale

When volume is too high for per-action approval, let the agent act and review a sample. The agent auto-categorized 1000 tickets today; a human reviews 5% randomly sampled, weighted toward low-confidence cases. Drift in the auto-categorizer shows up in the review rate quickly.

Async handoff: for escalation

Agent is mid-task, hits something it can't handle. Don't fail the session; hand off. Tier-1 → tier-2. Support agent → human rep. Billing agent → finance. The task pauses, joins a human queue, picks back up after resolution.

The confidence threshold

An agent can (and should) estimate its own confidence on any action: "I'm 95% sure this is the right refund" vs "I'm only 40% sure this is the right category." Route by threshold:

• >90% confidence → auto-execute (if the action is reversible or low-stakes).
• 50-90% confidence → queue for human approval.
• <50% confidence → return to user for clarification.

Track the calibration: when the agent says 90%, how often is it actually right? Adjust thresholds with real production data.

A worked example: customer refunds

• Auto-approve refunds under $50 with clear policy match (70% of cases).
• Pre-action approval for $50-$500 refunds (25% of cases; human clicks to execute).
• Async handoff to a human for refunds over $500 or policy-unclear cases (5% of cases).

Agent handles 95% of volume with a human touching 30% of dollars. One human reviewer can keep up with thousands of tickets a day because most cases are pre-approved low-risk.

Queue management

If agent output exceeds human review capacity, queues build up and the UX degrades. When you see this happening:

• Automate more at the top of the funnel (raise the auto-approve threshold for low-stakes cases).
• Hire more reviewers.
• Reduce agent throughput (slow the input rate).
• Tier the queue: urgent items jump the line, routine waits.

Pitfalls

• Human rubber-stamping. If reviewers just click approve on everything, the human-in-loop is theater. Design the UI so refusing is as easy as approving.
• No escalation path. Agent can't hand off, just fails. User experience collapses at the first edge case.
• Review fatigue. Too many requests kill attention quality. Prioritize which actions actually need a human.
• Queue SLAs not measured. If the average wait is 2 hours, users churn. Track and alert.
• Ambiguous confidence. Agent says "high confidence" on things it gets wrong. Calibrate using real outcomes.

What to do with this

• For each agent action, pick one of the three models (pre-approval, post-review, async handoff) based on stakes.
• Build the escalation path before the first customer needs it.
• Read safety + guardrails for the threats that motivate human-in-loop.
• Read agent handoffs for the agent-to-human transition details.