Match each HITL pattern to its defining tradeoff

**Human-in-the-loop** patterns decide when a person checks an agent's work. They exist because a fully autonomous agent can take real actions like sending money, deleting files, or emailing customers, and some of those actions cannot be undone. The first pattern is pre-action approval. The agent stops before every gated tool call and waits for a human to click approve. This is the safest option because nothing happens without a person seeing it. The cost is latency, the agent cannot move while it waits. The second pattern is post-action review. The agent runs the whole task on its own, and a human reads the logs afterward. This is the fastest option and gives the agent full autonomy. The danger is that a bad action has already happened by the time anyone looks. The third pattern is interrupt on threshold. The agent acts freely on routine, low-risk steps but pauses for a human only when an action crosses a risk or confidence threshold. It is the middle ground, fast on easy steps and careful on dangerous ones. The right pattern depends on how reversible each action is and how large the damage could be.

**Human-in-the-loop** patterns are three points on a curve that trades agent autonomy against safety and latency. The right pairing depends on how reversible an action is and how large its blast radius could be. Pre-action approval gives maximum safety. The agent pauses before each gated tool call and the loop stays blocked until a human responds. The cost is the highest latency, because every gated step waits on a person. You use it when actions are irreversible and high stakes. Post-action review gives full autonomy. The agent runs the entire task uninterrupted, and a human reviews the trace logs after the fact. It is the lowest-latency option. The risk is real, a harmful or irreversible action has already executed before anyone sees it. Interrupt on threshold sits between the two. The agent acts autonomously on routine steps and pauses only when a risk score or a confidence value crosses a configured line. It is more efficient than approving everything and safer than reviewing nothing. The hard part is calibrating the threshold so it catches the dangerous cases without nagging on every step.