This isn’t my experience. I find that, while the better models make fewer mistakes, a solid percentage of their recommendations and advice are actively harmful to the project. I was fairly “hands off” at the beginning of this project (so much fast progress!) but several months in, I realized what an absolute mess they had made. Every time they added a feature they often duplicated a large portion of the code, other agents did the same and this eventually resulted in an absolute spiderweb of dysfunction. It set me back a few months while i ripped it all out and came up with new processes to force structure, force code reuse and centralization. Just this morning, I caught an agent (grok) in another hallucination. I think that people get good results on smaller blocks of work, but I would worry that if you are fully hands off on a long running project with many cycles, where hallucinations can compound, then you might be surprised as to what they’ve actually written.