I feel like the exponential growth of the models will make this a non-issue in the future. So the tweaks we might make right now might not be even more with when even more powerful models in the future come.
Which is why I think additional intent preserving abstraction is where software coding agents are likely heading.