Is Loop Engineering New, or Repackaged Machine Learning? A Dog Classifier Explains

The article argues that loop engineering is largely an old idea, namely the basic machine learning training cycle and control-theory feedback loops, rebranded for AI agents. What is new is that the adjustable elements can now be text, files, or knowledge bases rather than only numeric parameters, and feedback can come as plain language. The author illustrates the concept by walking through how a dog-image classifier is trained, with a model, a loss function, gradients, and a stopping condition. The same structure is then mapped onto AI agents, where evaluation comes from compilers, tests, rubrics, or another model acting as judge. The piece says this pattern is useful for tasks that are too large, too open-ended, or unmonitored for a single conversation to handle reliably.