LLMs promise to build themselves, but reality is fragile. Despite years of research, fewer than one in ten agents actually utilize automated optimization. The issue stems from hidden design choices engineers must make manually. Before a model can learn from its own mistakes, someone must determine exactly which parts of the artifact are editable and which errors count as valid learning signals. These invisible decisions determine whether the system improves or fails completely. Researchers tested these factors across coding, games, and complex reasoning tasks. They found that starting conditions influence the solutions AI discovers, while how long a model analyzes past mistakes significantly impacts game performance. Furthermore, grouping errors to teach a model does not always yield better results; sometimes less data is optimal. The takeaway is clear: there is no universal method yet for configuring learning loops that work everywhere. This lack of a simple, standard approach remains the primary barrier to deploying these self-improving agents in real-world production.