Human-in-the-Loop

Human-in-the-loop means the agent does not act on its own for steps that matter. Before sending an email, spending money, deleting data, or posting publicly, the agent pauses and surfaces the proposed action to a person, who approves, edits, or rejects it. Then the agent continues from that decision. This is one of the most important reliability patterns in production. Models make mistakes, and some mistakes are expensive or irreversible. A human checkpoint at the risky steps catches errors before they cause damage while still letting the agent handle everything routine on its own. Implementing it well requires the framework to pause and resume cleanly, which ties human-in-the-loop to state persistence. LangGraph treats both as first-class: a run can stop at an approval node, persist its state, and resume later when the human responds. CrewAI, AutoGen, Google ADK, Semantic Kernel, and Mastra support it too. Lighter libraries leave it for you to build around the agent loop.