The first craft to emerge around language models was prompt engineering, the practice of wording a request so that the model returns what was intended. It is a real skill and it continues to matter, but it turns out to describe the smaller portion of the work. The moment a model is moved from answering a single question to driving a system that accomplishes real tasks, the difficult engineering is no longer in the prompt, it is in the loop that surrounds it, and that shift is where the discipline is now heading.
A single prompt is one exchange, a question posed and an answer read. Loop engineering concerns everything that happens when one answer is not the end of the process. The model produces a step, that step is executed, its result is fed back into the next prompt, and the cycle continues until the task is complete or the system decides to stop. Every genuinely capable system built on these models, an agent, a coding assistant, a research tool, is a loop of this kind rather than a single call, and its behavior is determined far more by the structure of the loop than by the wording of any individual prompt within it.
The consequential questions all live in the loop rather than the prompt. What condition terminates it, a completed task, a spending budget, a maximum number of attempts. What does the model see on each iteration, the full accumulated history or a summarized and trimmed context, a decision that governs both cost and coherence. What happens when a step fails, does the system retry, back off, escalate, or hand control to a person. How is the loop kept from cycling, repeating a failing action with undiminished confidence, a common and expensive failure mode. None of these are answered by better prompt wording, they are answered by control flow, state management, error handling, and explicit limits, which are the ordinary tools of building reliable systems applied to an unusually unpredictable component.
This is why the work is coming to resemble engineering more than writing. Prompt engineering asks what to say to the model, loop engineering asks how to construct a dependable system around something that is fluent, useful, and fundamentally non deterministic, and the second question is where reliability is actually won or lost. The prompt is one component among many, the loop is the architecture, and as these systems move from demonstrations into production the demand is shifting accordingly, toward the engineers who can build the structure that makes an unpredictable model behave predictably enough to depend on. That is the harder problem, and it is the one the field is only beginning to treat with the seriousness it requires.