The author expresses strong frustration with DSPy and GEPA, two tools aimed at improving LLM workflows. Despite the developers' intelligence and good intentions, the author finds their modular approach limiting. DSPy is designed to optimize AI tasks by allowing users to define desired outcomes without needing to understand the underlying complexities. For example, it enables users to create a search function that can be easily adapted to different models, reducing the need to start from scratch as circumstances change. GEPA, which stands for Genetic Pareto, is an optimizer that uses a genetic algorithm to enhance LLM programs. It evaluates potential prompt candidates based on their performance, keeping only those that excel in specific situations while discarding others that are less effective. Although GEPA shows promise in improving performance and efficiency compared to reinforcement learning methods, the author’s experience testing it on a multi-turn search task yielded disappointing results. They struggled with the output feeling disjointed and unnatural, akin to mismatched ingredients in a recipe. The author reflects on their initial excitement about GEPA but ultimately critiques its practical application. They aimed to optimize three key components of their search agent, but the outcome felt chaotic and ineffective. This experience highlights a disconnect between the theoretical benefits of these tools and their real-world usability. The author recognizes that while the concepts behind DSPy and GEPA are intriguing, their implementation can lead to frustration and unsatisfactory results.