Human-in-the-loop · May 26, 2026
The agent paused. The checkpoint fired. The approval request went to an inbox nobody checks. Multi-channel delivery is the missing half of HITL — here's how to route decisions to humans where they actually work.
Keep reading
Most reviewers ask "is this right?" before approving. The better question is "can this be undone?" The reversibility question reframes every review — it shifts the focus from correctness to recoverability, from immediate fit to downstream survivability. A reviewer who asks "can this be undone?" first catches more failures, makes better decisions, and produces more durable audit trails. Here is why this one question change makes HITL work.
Every HITL review is a forecast — the reviewer predicts the action's outcome. Most reviewers forecast by asking "will this work?" A small minority ask the inverse: "how will this fail?" That minority is right more often than the majority. The pre-mortem — imagining the failure before approving the action — is the cheapest, most effective single intervention HITL systems can adopt. Here is why, how to design for it, and what it changes.
Send the same action to the same reviewer at 9am, 11am, 2pm, and 4pm — you will get four different decisions. The action is identical. The reviewer is identical. The decision differs. This is the judgment gradient: the same reviewer applies different judgments at different points in their session. The gradient is the largest source of inconsistency in HITL, and the most invisible. Here is why it happens, what it costs, and how to design for consistency.