Parallelize yourself, not your agents

Is it arguing that we need to supervise better than building systems that enable agents to make better decisions?

I would argue that unique decisions should only be made once (I'm building it). It is a gated mechanism, where confidence is 0<x<y<1 >y - decision delegated from x to y - rapid decision, upto 5 lines of context and simplified ask <x - full context for elaborate decision

Everything <y is fed back for making decisions. Plus, decisions are logged with fragments/context used.

By Amdahl's law, you'd have to automate 99.9% of decision-making to reach 1000x, 100x is at 99%. That's near impossible unless y is closer to 0, i.e., near unbound delegation.