Agents need the ability to code but also to objectively and accurately evaluate whether changes resulted in real improvements. This requires skills with metrics and statistics. If they can make those reliable then self-improvement is basically assured, on a long enough timeline.