human-in-the-loop

name	human-in-the-loop
description	Design where, when, and how a human gates, reviews, or rescues an LLM agent — without turning the agent into a button labelled "approve". Use when the user is building an agent that takes irreversible actions or operates in regulated workflows and mentions human-in-the-loop, HITL, approval gate, escalation, review queue, oversight, or asks "when should a human approve this?" / "how do I add review without killing the agent's speed?".
tags	["safety","oversight","production","workflow"]

Human-in-the-Loop

A human gate is a strong control — and an expensive one. Place it where the cost matches the risk.

The default mistakes are equal and opposite: gate everything (the agent is a slow button), or gate nothing (the agent ships an irreversible mistake). Design the gates on a per-action basis, by failure cost, with the human's time treated as a scarce resource.

When to use this skill

• The user is building an agent that takes real-world actions (money, messages, data changes).
• The user is in a regulated workflow (healthcare, legal, finance, HR).
• The user reports "the agent did something bad once" and is reaching for blanket approval gates.
• The user reports "approval gates everywhere are killing throughput".

Tier the actions

Not all actions need humans. Classify before designing.

• Auto-execute. Read-safe, low-impact, reversible. No gate.
• Notify. Action executes; human is informed (digest, dashboard, audit log). Catch errors via sampling.
• Confirm. Action proposed; human approves before execution. Synchronous gate; cost is real.
• Review. Action executes; human reviews a sample asynchronously. Catches systemic issues without blocking.