A collaboration pattern where a human actively reviews and redirects an agent during a session.
Human-in-the-loop is a working pattern where one or more humans pair with an AI agent during a session, reviewing its actions, redirecting its approach, or collaborating in real time. The human is present and engaged throughout the work, not simply approving or rejecting individual tool calls in isolation. The agent's full output — reasoning, file changes, tool sequences — is visible to the human as it happens.
The mechanism varies by implementation: some setups use a shared terminal session where the human watches the agent work and can interrupt mid-turn; others use a UI that surfaces agent actions for review before they execute. The defining characteristic is real-time human engagement rather than deferred approval.
HITL is most valuable for high-stakes or irreversible actions — schema migrations, production deployments, content deletions — where the cost of an error exceeds the time saved by full automation. It introduces friction and reduces raw throughput, so teams weigh the risk profile of each task against the overhead of continuous human attention.
Open questions include how to design HITL interfaces that give humans meaningful situational awareness without overwhelming them with detail, and whether HITL patterns can be systematically applied to reduce catastrophic errors without creating human bottleneck points that defeat the purpose of automation.