Interesting. The review bottleneck matches what we’ve seen too: agents can ship a lot of audio/UI code fast, but the hard part is knowing whether the scheduler, sample playback, and MIDI path are actually correct without listening and clicking through. Spec + adversarial PR review is a solid setup. Curious how you’re thinking about MIDI next — keeping generated ideas as editable notes/clips, versus baking more into rendered audio? We’re biased toward SoundFonts + editable MIDI for that reason.