Surprising to me that the major agent models are misaligned in such different ways. Codex will sit and spin, deep frying 1 micro feature with infinity tests and guards and defensive engineering, but does not care about the ultimate goal and makes no progress toward it. Claude just skips ahead, stringing out One placeholder after another until it can claim the whole project is finished. They clearly are tricking different types of judges in training. Both can realize and recognize what that they are doing this if you point it out, but will Go right back to it as soon as they think you’re not looking
· 2 min read
Archives