Measuring AI agent autonomy in practice