Fwiw, top of my own pain-point list (I suppose given the first item, that's a pun) includes:

External/policy-based throttling for temperature control. Unthrottled, my laptop bottom goes skin-burn hot. But fixed compute caps can have non-linearly dreadful performance impacts in particular cases. Plan is a runtime knob, to replace manual limits-kludgery.

I'll use models which barely fit in VRAM+RAM, and are order-1 tok/s slow. So tool call step overhead can be painful - a world where `ls` costs tens of seconds. Plan is blending harness plugins with inference loop, for "no, don't stop - I already have the call result for you - just keep going" (and also some logit games).