The agent review loop

An AI workflow is only safe if a human can supervise it without re-doing the work. The agent review loop is the smallest structure that makes that true.

The loop

1. The agent prepares a draft with its sources attached.
2. The human gets one screen: the draft, the context, and three buttons — approve, edit, reject.
3. Every action writes an audit entry. Nothing self-publishes.

The discipline

Confidence is shown, never enforced. The human is always the last gate. Build it this way once and every future agent inherits a review surface you already trust.