Human in the Loop

The engineering requirement behind the phrase is that the run must be interruptible and resumable. An agent that has to hold an open connection while waiting for approval cannot wait long; one whose state is checkpointed can pause for hours, survive a deploy and continue from exactly where it stopped. This is why durable execution and graph-based agent runtimes treat interruption as a first-class primitive rather than an afterthought.

Where the checkpoint goes is a risk decision, not a stylistic one. Reversible, low-consequence actions should proceed unattended — gating them trains reviewers to click approve without reading, which is worse than no gate at all. Irreversible, expensive or externally visible actions — sending communications, moving money, deleting records, publishing — are where a real review earns its latency. The reviewer also needs enough context to judge: what the agent intends, why, and what evidence it relied on.

The misconception is that a human in the loop is a temporary scaffold to be removed once the model improves. In domains with regulatory, financial or reputational exposure, the review point is the control that makes automation acceptable at all, and it stays. What legitimately shrinks over time is its scope — narrowing which categories require approval as evidence accumulates, rather than removing the mechanism.