Human in the loop
A workflow pattern where the agent pauses for human approval at one or more checkpoints before continuing.
Human-in-the-loop (HITL) is the safety mechanism that makes semi-autonomous agents work. The agent does the busywork autonomously, then stops at a gate: "I'm about to send this email — approve?".
Done well, HITL gates land at the points where the cost of being wrong is asymmetric — sending a message, deploying code, paying an invoice. Done badly, gates land everywhere and turn the agent back into a slow assistant.
A common pattern in 2026: start every workflow at "gate everything", remove gates one type at a time as trust accumulates, and keep gates permanently on anything touching external relationships or money.
Frequently asked
Does human-in-the-loop slow down agents too much?+
Only if you gate every step. Well-designed HITL gates only the irreversible actions — usually less than 5% of total steps — so the throughput cost is small but the safety win is large.