Human-in-the-loop is an architecture, not a checkbox

Every agent vendor claims human-in-the-loop. Usually it means a confirmation dialog. Real HITL is an architecture question: where in the pipeline do human judgments create the most safety per second of human attention? Get it wrong and you've built a system where a person babysits every step, automation that saves no time.

The pattern that works: agents do the work end-to-end, then present completed work for review, approve, don't type. The operator sees what the agent prepared (a chart, a booking, an extracted document) as a reviewable diff against reality, approves or corrects, and the write lands. The human supplies judgment; the agent supplies labor. Reversed roles is how you get either rubber-stamping or burnout.

Placement is everything. Reviews belong at write boundaries, the moments work becomes irreversible, not scattered through the pipeline. Everything before the boundary can run fully autonomous because it's recoverable; everything at the boundary gets a human gate until the eval history earns it autonomy. Trust is granted per action type, backed by data, not per vibes.

The unglamorous part that decides success: the review surface itself. I ship a desktop app for this, native, fast, keyboard-driven, because operators live in it for hours. Batched approvals, sensible defaults, one-keystroke corrections. If reviewing an agent's work is slower than doing the work, your automation has negative value and your operators know it before your dashboard does.

Metric that matters: seconds of human attention per completed unit of work, tracked over time. It should fall as evals accumulate and trust graduates. If it doesn't, you built a very expensive suggestion engine.