Human-in-the-Loop (HITL)
A design pattern where AI agents pause and ask a human to approve or correct high-stakes actions.
HITL is the most reliable safety pattern for agentic AI in production. Instead of letting an agent run end-to-end, you insert checkpoints it drafts the email, you approve sending; it proposes a database change, you approve running it.
The trick is choosing the right checkpoints. Too many and the agent is just an autocomplete; too few and you've shipped something dangerous. Most production agents check in at irreversible actions (sends, payments, deletes) and at confidence-uncertain steps.
Good HITL UX makes it easy to approve and easy to correct. The agent shows what it's about to do, you click yes/no/edit, it proceeds. Done well, you get most of the speedup with little of the risk.