tl;dr

> This is driven by two developments: more parallel agents and more work done before human feedback is needed.