Human-in-the-loop

Fully autonomous agents are a small fraction of production systems. Most agents in production have humans in the loop at specific decision points. Knowing when to pause for human review is a design decision, not a limitation.

When to pause for human

The interaction model

Pre-action approval

Agent proposes action. Human sees it, approves or edits. Then it executes.

Post-action review

Agent takes action. Human reviews later. For less-critical actions where speed matters.

Async handoff

Agent pauses task, surfaces to human queue. Human resumes with annotations. Agent continues.

Queue management

If agent actions exceed human review capacity, queue builds up and UX degrades. Either:

The confidence threshold

Agent can estimate its own confidence. Route high-confidence actions through automatically; queue low-confidence for review. Over time, track which actions were approved vs changed, calibrate the threshold.

Agent proposing vs executing

Strong pattern: agent always proposes, human clicks to execute. Keeps human as the actor of record. Lowers liability and improves trust.