I like having different models review each other—for instance, having Claude write code for Codex to review, or having Codex write it and Claude review it.
Not necessarily; it depends on cost and stability. You don't need to develop everything yourself—in most cases, using existing solutions is more reliable.