I've asked Codex with GPT 6 Astra to *review" a one time benchmarking script for any mistakes (built by Claude Code using Opus 5.5) and it refactored the shit out of it claiming all sorts of stuff without even asking about the context in which it was developed.

If I was to employ them to review the code without giving each the same baseline multi-page prompt, they go into endless loop of "improvement" with no end goal in sight.

More and more frequently, I instruct frontier models to stop and go back to the task at hand.