Glossary
What is human in the loop?
Human in the loop (HITL) is a workflow design where a person must approve or reject a specific step before an agent's output proceeds, usually gating irreversible or customer-facing actions. The checkpoint is fixed by the system that orchestrates the agent, not the model itself, and the human only reviews what that orchestrator flags.
Why it matters
In the loop means the human's decision blocks the process until they respond. On the loop means the human can intervene, but the agent proceeds by default if nobody does. Most production agent systems use both: routine steps run on-the-loop, high-stakes ones (a deploy, a payment, an email sent under your name) run in-the-loop. The question that matters is who decides which is which. Not the model. You don't ask an LLM whether a step needs human review and trust the answer, because the model has nothing at stake in being wrong. The orchestrator, the code that wraps the agent, decides in advance: this action requires approval, that one doesn't. That's a design decision made once, not a runtime negotiation the agent talks itself out of. For agent loops that ship UI, the checkpoint that's easiest to skip is "does this actually look and work right for a real user." That's not a call an agent can make about its own output. It has to be a human, at a defined point, looking at the real thing.
How Patchrooms uses it
Patchrooms is the in-the-loop checkpoint for AI-built apps: review a live preview instead of a diff. One click on a broken element becomes a structured report (selector, screenshot, page URL, comment) that an agent loop reads back and fixes. The checkpoint stays a single, cheap action on your side. The agent still does the fixing.
Try the live demo