Who actually broke your AI agent: the model or the wrapper?

When AI agents fail in production, developers often debate the cause: is it the language model, a flawed prompt, or an API issue? A new study from Scale AI offers a systematic approach to diagnosing these failures. The authors present a taxonomy of 41 failure modes that helps pinpoint exactly where a system breaks down. Analysis shows that 'model failure' is rarely solved by simple fine-tuning. In practice, most issues lie in the 'wrapper'—the prompt architecture, tool-calling logic, or data handling. The article examines four real-world cases, demonstrating that swapping the model is not a universal solution. This material is valuable for engineers working on AI reliability, as it shifts the focus from endless model switching to debugging the infrastructure and the logic of how an agent interacts with the outside world.