Sure, high level is the goal. But the question is whether the abstractions are the correct ones that fit the problem. Almost all software that I have encountered that was painful to work with chose a framework that did not apply to their situation.

E.g., develop a generic user interface framework which makes it very quick to produce a standard page with a series of standard fields but at the same time makes it very painful to produce a non-standard layout. After that is done it is 'discovered' that almost all pages are non-standard. But that 'discovery' could also have been made in five minutes by talking to any of the people already working for the company....

Another example: use an agent system where lots of agents do almost nothing, maybe translate one enum value to another enum value of another enum type. Then discover that you get performance problems because agent traffic is quite expensive. At the same time typical java endless typing occurs because of the enormous amount of agent boilerplate. Also the agents that actually do something useful become god classes because basically all non-trivial logic goes there....

> Sure, high level is the goal. But the question is whether the abstractions are the correct ones that fit the problem.

Not quite. The path to high level always involves abstractions that fit the problem. There is still room for a decision to replace high-level with low-level in some very specific bits of a hot path, but that decision also takes into consideration the tradeoffs of foregoing straight-forward high-level solutions with low-level versions that are harder to maintain. The sales pitch to push code that is harder to maintain requires a case that goes way beyond performance arguments.