Loop Engineering, Chapter 6: Designing Human-in-the-Loop Controls

This is Part 6 of a series walking through my book Loop Engineering: Scaling and Governing Agentic AI. In the previous chapter, we built deterministic governance. This one is about the decisions that still need a person.

The phrase “human in the loop” gets used as though it were a single setting, on or off. It is not. The chapter starts by replacing the binary with a ladder of four rungs. At Assist, the agent drafts and the human does the work. At Augment, the agent works and the human reviews every action. At Automate, the agent acts on its own within a known envelope and escalates only the exceptions. At Agentic, the agent works and the human audits outcomes after the fact. Each rung withdraws human effort a little further, and the discipline is to climb it deliberately, on evidence, rather than drift up it because the volume left no choice.

The reason the ladder matters is that human review is not free and not reliable. As an agent gets good, the human reviewing it sees a longer and longer run of correct actions, and approval becomes a reflex. This is automation bias, and you cannot fix it by telling people to stay vigilant — vigilance is a finite resource that automation steadily depletes.

Designing against the rubber stamp

The fixes are structural. A decision cap limits how many approvals one person issues before the queue blocks, converting an endless scroll into a bounded, attention-worthy batch. Effort symmetry makes approving and rejecting cost about the same, so the path of least resistance is not always “yes.” Blind sampling slips known cases into the review stream to measure whether the reviewer is actually reading. None of these makes oversight cheap; they make it real.

The back half of the chapter is a catalog of interface patterns, each matched to a level of risk. Atomic confirmations are high-friction modal dialogs reserved for discrete, irreversible acts. Session scopes grant bounded autonomy up front — read-only versus read-write — so the agent does not interrupt for every step. Interactive parameter editing lets a human fix the one wrong field in a proposed action instead of choosing between blanket approve and blanket deny. And scale-aware impact previews translate an abstract intent into a human-readable consequence — “this will modify 2,315 records” — before anyone clicks yes, which is how you stop a bulk-destruction mistake while it is still a preview.

Tomorrow: human review cannot scale to the volume agents produce. Chapter 7 opens the final part with the discipline that can — automated evaluation.