Describe three patterns for inserting human approval into an agent workflow and their tradeoffs

A **human in the loop** pattern is a checkpoint where a person can inspect or approve what an agent is about to do, or has already done. The point is to keep a human in control of the risky parts without supervising every harmless step. The first pattern is approve before execute. The agent pauses before a tool call and waits for a person to click approve. It is the safest option because nothing risky happens without sign-off, but it is the slowest because the agent sits idle waiting. The second pattern is edit and continue. The agent proposes an action or a draft, the human edits it, and then the corrected version runs. This keeps the human in control while still letting the agent do the heavy lifting of drafting. The third pattern is post-hoc review with an undo window. The agent runs fast on its own, every action is logged, and a person reviews afterward. It is the fastest, but it only works when actions can be undone in time. A fourth, increasingly common pattern is escalate on low confidence. The agent runs alone but asks for help only when it is unsure. The key rule across all of them is to choose based on how easy it is to undo a mistake.

**Human-in-the-loop** patterns insert a person at one of three timeline positions, and the right one depends on action reversibility, not preference. Approve before execute pauses the agent before a tool call and blocks until a human confirms. It is the safest pattern because no irreversible action runs without sign-off. The cost is latency: the loop is idle while the human responds. Edit and continue is a softer variant where the human can amend the proposed action before it runs, which keeps autonomy higher. Post-hoc review with an undo window lets the agent run at full speed and logs every action for asynchronous review. Latency is lowest, but the risk is real. An irreversible action like a sent payment may have already happened by review time. It only works when an undo window covers the gap, such as a delayed-send queue or a soft delete. Escalate-on-low-confidence runs autonomously but interrupts only when a confidence or risk score crosses a threshold. It balances speed and control, so cheap actions proceed freely while risky ones pause. The catch is calibration: a bad threshold either floods the human or lets bad actions through.