AI coding agents can execute shell commands. what’s the safest way to control them in production?