What is Human in the loop?

Human in the loop is a design pattern where an AI system pauses for human review rather than acting fully autonomously. Instead of letting an agent execute every step end to end, the workflow surfaces certain decisions, sending an email, deleting records, spending money, merging code, to a person who approves, edits, or rejects them before they take effect. HITL is the practical answer to the fact that models hallucinate, misread intent, and can be manipulated by prompt injection or tool poisoning: a confirmation gate on irreversible actions turns a silent failure into a caught one. It also generates training and audit signal, the human's corrections can be logged for evaluation or future fine-tuning, and the approval trail supports compliance. In the Model Context Protocol this maps directly onto elicitation, where a server can ask the client to gather input or confirmation from the user mid-operation, and onto client UX that shows the user which tool an agent wants to call and what arguments it will pass. The engineering trade-off is autonomy versus safety and throughput: gate too much and you lose the speed that made automation worthwhile; gate too little and a single bad action slips through. Mature systems calibrate this by risk, auto-running low-stakes, reversible steps and reserving human review for the consequential ones.