Where in an agent workflow should a human approval gate be placed for an irreversible action?

A **human in the loop** gate is a checkpoint where the agent pauses and waits for a person to approve an action before continuing. The whole question is where to put that checkpoint. The answer for an irreversible action is: before the action runs. Some actions cannot be taken back once they happen. Sending an email, deleting a database record, charging a credit card, deploying code to production. If you let the agent do these and review afterward, the harm is already done by the time the human looks. A pre-execution gate works differently. The agent decides it wants to send the email, then it stops and shows the human exactly what it is about to do. The human approves or rejects. Only an approval lets the action proceed. The other placements all fail for irreversible work. Reviewing after execution is too late. Gating only on failure misses successful but wrong actions. Approving once at the very start cannot cover an action the agent has not decided on yet. The simple rule: if you cannot undo it, approve it before it happens, not after.

For an irreversible action, the human approval gate must sit before the action executes. This is the only placement that lets a person actually prevent the action. The reasoning is timing. An irreversible action cannot be undone once it runs. A sent email, a deleted record, a charged payment, a production deploy. A post-execution review can describe what happened, but it cannot reverse it, so the gate arrives too late to help. A pre-execution gate interrupts the loop at the moment the agent proposes the tool call. The runtime pauses, surfaces the proposed action and its arguments to a human, and resumes only on explicit approval. A rejection blocks the call entirely. The other options fail on the same axis. Gating only on failure misses the dangerous case, which is a successful but wrong action. Approving once at the start of the workflow cannot cover actions the agent has not yet chosen, since agents decide each step from the latest observation. The deciding axis is reversibility times impact. Gate destructive or high-cost actions before they run. Auto-approve cheap reversible ones so the human is not flooded with trivial confirmations.