Where the Approval Goes

You have four places to put a human check, and they have very different costs and very different safety properties.

A. BEFORE THE AGENT STARTS

Human approves the goal, then the agent runs freely.

Cheap. Catches "wrong task" but nothing after that.

B. BEFORE EACH IRREVERSIBLE ACTION

Agent works, pauses for a yes on the risky steps.

The default for most products. Good safety per unit

of friction.

C. BEFORE EVERY ACTION

Human confirms each step.

Very safe. Also slower than doing it manually, so

users will route around it. Rarely correct.

D. AFTER, WITH A WINDOW

Agent acts; human can reverse within a set period.

Fast and fine for reversible things only.

Most products want B for irreversible actions and D for

reversible ones. That combination gives you most of the

safety for a fraction of the friction.