Is automatic gain compensation for EQ something we want?

[MirkoVanHauten]

Actually, I even think it's not a good idea: If I want to turn down something I don't want the rest to go higher.

Exactly. Just imagine mixing stuff, eliminating some unwanted frequencies and suddenly the element gets louder and influences all the following effects. Sounds rather like a nightmare and source for neverending back & forth.

[Dasheesh]

but, you don't HAVE to do it. It's optional. It's a feature. what AGR does is take the signal to the point of distortion. anything beyond that point is coloring the sound. you are free to gain after that point.

[Niowiad]

To me it's quite straightforward, really.
And it goes for most other sound processors (compression, distortion, etc...) featuring AGC.
But specifically talking about EQ:

Whenever listening to an element "in isolation", AGC can be a very useful feature to concentrate on the frequency balance while minimizing the effect loudness has on the overall perception of that element after boosting/cutting frequencies.

Whenever listening to an element "in context", that's when AGC likely becomes counterproductive.
If you're targeting a specific frequency range of an element in the mix, according to what you're hearing in context, you obviously want to affect that range and nothing else.

Now, the OP asked about mastering, which falls under the "in isolation" case.

[cron]

It doesn't make sense to me in EQ, or indeed any non-full spectrum process, at all. I'm reminded of an ancient thread I commented on where somebody couldn't understand why playing 2 notes at the same time returned higher loudness levels than one note in isolation because both sounded equal in loudness to him. Also a more recent thread where someone couldn't understand how dynamic range was reduced when you set a compressor to allow transients to poke through. Both of those threads highlighted the core problem: you can throw all the psychoacoustic weighting in the world at the wall, but the kinds of algorithms used in audio measurement bear so little relation to how we actually hear that they're IMO next to useless when it comes to anything other than comparing one wideband signal to another, similar wideband signal - especially when time is short. I've only come across an equal loudness feature outside these parameters in Slick EQ, and it's immediately obvious when using it that equal perceptual loudness isn't maintained, even when making small adjustments.

[heavymetalmixer]

It doesn't make sense to me in EQ, or indeed any non-full spectrum process, at all. I'm reminded of an ancient thread I commented on where somebody couldn't understand why playing 2 notes at the same time returned higher loudness levels than one note in isolation because both sounded equal in loudness to him. Also a more recent thread where someone couldn't understand how dynamic range was reduced when you set a compressor to allow transients to poke through. Both of those threads highlighted the core problem: you can throw all the psychoacoustic weighting in the world at the wall, but the kinds of algorithms used in audio measurement bear so little relation to how we actually hear that they're IMO next to useless when it comes to anything other than comparing one wideband signal to another, similar wideband signal - especially when time is short. I've only come across an equal loudness feature outside these parameters in Slick EQ, and it's immediately obvious when using it that equal perceptual loudness isn't maintained, even when making small adjustments.

Two notes about SlickEQ:
1) It can be used as a saturation plugin alone.

2) The saturation has an auto-gain feature separated from the EQ and it can't be turned off.