I've found detailed logging plus approval gates for high-risk actions helps—lets agents move fast on reversible stuff, slows them down where it matters.