Some teams I know started moving all linting and validation into LLMs and markdown files, the reasoning was that they asked agent to review changes that were later found to be buggy and agents flagged these bugs. And that’s great but my immediate worry was that: it finds it now, but will you check every new model for all these bugs? And what if you need to switch providers? Or your provider silently tweaks the model to improve their margins? The answer was “well also build deterministic linters for all these problems”, but you know that’s just wishful thinking - if you can get 99.9% done with just a little bit of fuzzy text and the next 0.1% is a lot of work it won’t happen. So you have model behaviour lock-in that will bite you suddenly when you least expect it. Or maybe LLMs will just keep getting better and cheaper? It’s all a casino right now.