Well, this is not "so" hard about it.
UAE have it.
Minimig/Mist/Replay have it.
Vampire AGA is work in progress.
There were also other implementations never released.
All those implementations are differents and works rather good.
For example, Pinball Illusion is running quite perfect on a AGA Vampire summer 2017 core.
RTG is easier, yes, but also faster.
That said, with chipram mapped to fastram like there was on some proof of concept shown that AGA can be super fast. On that core, Metropolice demo was running even better than in RTG.
The difficulty is always same, the last 5% compatibility issues can need infinite work to find the perfect implementation, this because AGA was never 100% documented. And this takes also some room in a fpga.