For people building tools that AI agents use through the command line: what design choices ended up mattering in practice?