What is the human-in-the-loop (HITL) pattern?
LLM & AILevel: IntermediateSafety & Reliability
What is the human-in-the-loop (HITL) pattern?
Hints
- Identify the stage: pretraining, adaptation, generation or evaluation.
- Name the inputs, constraints and a case where the approach fails.
Answer
HITL pauses agent execution at certain checkpoints to request human input or approval:
Agent: "I'm about to send this email to 10,000 customers. Please review: [preview]"
Human: [approves / edits / cancels]
Agent: [proceeds or aborts based on human response]
When to apply HITL:
- Irreversible actions (send, publish, delete, execute in production)
- High-stakes decisions (financial transactions, medical decisions)
- Confidence below threshold — agent explicitly flags uncertainty
- Novel situations outside the agent's training distribution
UX consideration: too many checkpoints defeats automation's purpose. Use risk scoring to calibrate HITL frequency; start conservative, loosen based on trust data.
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0