Prompt Engineering vs Loop Engineering vs Graph Engineering: What Changes at Each Layer

The article explains that prompt engineering, loop engineering, and graph engineering are not competing techniques but stacked layers of control. Prompt engineering handles a single model response and assumes human iteration. Loop engineering adds a cycle of observe-act-verify-recover, making agent behavior programmable. Graph engineering organizes multiple agents via two simultaneous graphs: a stable org graph (who owns what) and an ephemeral work graph (what to do now). The article provides guidance on choosing the layer: if a person reads every output, prompt is sufficient; if "done" can be checked mechanically, a loop can be built; if independent branches must run concurrently, a graph is needed. Key takeaways include that a loop is only as good as its stop condition, and most tasks never need the top layer.