Definition
Human-in-the-loop (HITL) is an AI agent design pattern where a human is included at a defined point in an automated workflow to review, approve, modify, or reject an output before the process continues. The human does not replace the automation. The automation runs up to the decision point, stops, and waits for human input before proceeding.
How It Works in Practice
A well-designed HITL workflow is not a checkpoint applied to every output. It is a deliberate pause placed at decisions that carry real consequences, where human judgment changes the outcome in ways the AI cannot reliably replicate on its own.
The workflow runs fully until it reaches the defined gate. The AI surfaces the output and the context the reviewer needs to make the decision. The human approves, rejects, or edits. The agentic system resumes from exactly that point, with state preserved, without restarting the workflow from the beginning.
A quoting workflow is a common example. The AI handles configuration logic, compatibility checks, and pricing rules. A human reviews the output and decides whether to send the quote. The AI removes the bottleneck of manual processing. The human owns the decision that carries relationship and financial risk.
Why State Management Is the Hard Part
The engineering challenge in HITL is not the design logic. It is state management. When a workflow pauses for a human decision, whether for thirty seconds or three days, the system must hold every prior decision and piece of processed data in place. When the human acts, the workflow resumes cleanly from that exact point without losing context.
Most standard automation tools are designed for sequences that run to completion and do not handle this natively. Delegation between agents and memory in agents are closely related concepts that inform how a production HITL system preserves context across an interruption.
Wippy is built to handle this natively. When a workflow hits a review gate, Wippy holds it in a waiting state, routes the decision to the right person with the relevant context, and resumes from exactly where it paused. Nothing restarts. Nothing is lost.
Common Failure Modes
A review gate that requires the reviewer to understand the full workflow history before making a decision is a bottleneck, not a safety feature. If the reviewer spends more time reconstructing context than evaluating the output, the gate is surfacing the wrong information. If rejections send the entire workflow back to the beginning, the state layer was not designed for HITL from the start.
