I’m not sure I buy this. GPT-5.2 codex still makes design errors that I as an engineer have to watch and correct. The only way I know how to catch it and then steer the model towards a correction is to be able to read the code and write some code into the prompt. So one can’t abstract programming language away through an agent…