WTF is Human-in-the-loop?
Pausing an agent so a person can approve, correct or reject a step before it continues.
Also called: hitl, human approval
TL;DR
Let the agent do the tedious parts, but make it ask before anything risky: sending an email, deleting files, spending money. It is the simplest and most effective safety control for real-world agents.
How it works
Implemented as approval gates in the agentic loop: certain tool calls are marked as requiring confirmation, the run pauses (often persisted so it can resume later), and the human's decision is fed back as a tool result. Design questions: which actions need approval, how to show enough context for a fast decision, and avoiding approval fatigue where people click "yes" without reading.
Origin
- Coined / introduced by
- Long-standing automation concept
- Story
- Used in control systems and machine learning long before LLM agents; now a standard agent safety pattern.