I think "operate at a higher level of abstraction" is sometimes backwards. The problem with legacy systems is that the low-level details are the requirements.

I think "operate at a higher level of abstraction" is sometimes backwards. The problem with legacy systems is that the low-level details are the requirements. Things that look like implementation details often encode 20 years of business rules. An AI agent doesn't know which weird-looking piece of code is accidental complexity and which weird-looking piece of code is keeping a $10M customer from breaking. So I don't think the answer is necessarily to give the agent higher-level instructions. Sometimes the answer is to give it less autonomy.