That’s a funny but un-generous take on the approach, because it’s actually quite effective. There’s a guy on the Fediverse who built a local model that self-governs this way; it’s kind of like peer review. It also gets rid of all the sycophancy and “how about I also do this?” pleading of commercial agentic AI frameworks.

You still need to be in control as the user – this doesn’t make the AI fault proof in terms of misinterpretation, overengineering, etc.