One thing that stood out is the distinction between fixing the output and fixing the producer. That's a mindset many teams miss with coding agents. If the same issue keeps resurfacing, the solution isn't a longer prompt it's deciding whether that lesson belongs in repository instructions, task-specific skills, deterministic checks, or regression tests. Treating recurring failures as process improvements instead of one-off corrections is what makes an agent workflow actually mature over time. Nice breakdown of where each type of knowledge should live.