Yes, one of the obvious ways to me to use these models is to tell them to write such a program. It can then go figure out data extraction and normalization. This is "the harness improving itself". Have it write tools to do its tasks.

I think you missing the problem with composition itself. The tool generation is the easy part. Knowing which tool is needed, specifying it correctly, validating it against current context, knowing why it failed, and deciding when it needs replacement are separate tasks, and they compose, and composition compounds and fails.

I may not be personally familiar with the latest and greatest moving target in agentic AI, but I am familiar with AI-generated "slop code". What I see is consistently defective, to the point where I would not trust an AI-generated data pipeline to produce remotely accurate results, even though messing up a basic data pipeline that much is a difficult task for humans. (Then there's the tendency for an LLM post-processing the tool output to smooth over such flaws, massaging or fabricating conspicuously missing or corrupt data to hide what would normally be extremely obvious warning signs.) If it is, as the proponents claim, even possible to entice these systems to produce decent programs (other than by direct plagiarism), it must require skill way beyond that of the average prompter with 18 months' experience to explain my observations.

When I say "write an ordinary computer program". I mean just writing the program, in a programming language. Your comment expresses disagreement with me, so should have started with "No", not "Yes".