TOPIC 04 OF 06
Keeping a checkpoint
The human 'are you sure?'An approval gate pauses the agent before a consequential action — sending, paying, deleting, posting — and asks you to confirm. It's the single most valuable safeguard, because it stops a mistaken or hijacked action from becoming an irreversible one. You decide which actions are 'gated' and which the agent may just do.
"Draft the replies and show me before sending" gates the send. If the agent read a booby-trapped page that told it to email your files to a stranger, that instruction now dies at your confirmation screen instead of executing silently.
A gate only works if you can actually see and refuse the action — a rubber-stamp you always click 'yes' on is no gate at all. Reserve gates for what genuinely matters so they stay meaningful.
Pick the one action you'd most regret an agent doing wrong. Make that the thing it must always ask you about.