My point is the code for feedback loops is architecturally simple: when reading is wrong correct.

By boosting the accuracy and frequency of the reading you can get pretty good results.

But that has little to do with LLMs, or LLM generated code.