Every Retrieval-Augmented Generation (RAG) system is built upon three engineering layers stacked on a single LLM call: prompt, context, and loop. Prompt engineering defines the call itself, including system messages, instructions, and output schemas. Context engineering manages what fills the model's finite window, involving retrieval, compression, and deciding what to exclude.

Loop engineering governs the process around the call, determining when subsequent calls are triggered, when the system stops, and how it recovers from failures. Understanding these layers is crucial for building and debugging RAG systems. While a narrative of sequential evolution from prompt to context to loop engineering exists, the underlying patterns predated the terminology, with all three layers coexisting from the start.

The bottleneck in RAG development has shifted over time. Initially, prompt quality was the primary constraint. As models improved, the focus moved to context management, and more recently, to loop engineering as systems handle longer, more complex interactions. This shift is driven by advancements in model capabilities, increasing context window sizes, and the evolution of use cases from single-shot Q&A to multi-turn, agent-driven processes. Each layer addresses a specific challenge, with prompt engineering concerning the model's input for a single call.