This is a well-articulated point. In my experience building AI systems, the cognitive load issue shows up most when you're juggling model logic, API integrations, and deployment pipelines all at once , the architecture starts feeling more like a liability than an asset. The abstraction layer suggestion is spot on. Clean separation of concerns genuinely reduces mental overhead and makes the system easier to reason about long-term.