Bacteria are "simple" enough that we can just append some DNA sequences and have them develop stuff we want. It's predictable enough that people already created "standard library" of DNA blocks - so people can work with bacteria in a kind of dataflow environment, with underlying DNA representation hidden. You connect a "molecule sensor" block with an "amplifier" block, an "inverter" block and a "light up" block, and you get bacteria that start to glow in presence of some metal, etc.

But of course, once we move to the level of multicellular organisms, the complexity explodes. Here, as you say, we basically know that something in an unaltered system sorta correlates with a particular phenotype and if we change it then maybe it won't blow up.

The compiler is called "genetic engineering" and it's a 30 years old field. In particular genetic engineering allows you to knockout specific genes and observe the phenotype. This technique is commonly used on bacteria, plants and animals. There is an ongoing mouse knockout project, an attempt to study every mouse gene by turning it off. It is 50% complete so far. By the way, the experiments are performed mostly by hand.

>We're very far away from understanding how a particular DNA string relates to phenotypes.

This problem can be solved much faster with a large scale effort. Thousands of automated experiments can run in parallel while recording the phenotype of the animals (including results of behavioral tests) and storing it for later analysis. On this genotype->phenotype dataset one can train an ML model. Actually there is a company that is trying this approach (just ML part, without automated experiments) http://www.deepgenomics.com/

I doubt that one company will be enough, though. With serious funding there could be much more progress in this area. We could have a full knockout map of mouse genome in less than a decade if we really wanted.