I overpowered AI by inventing "Brilliant" code by AI opinion itself

As it always happens to me, I started from describing a simple API client, then I started playing around Mixins, and at this point I really wanted to make my own Mixins library that is just faster than ts-mixer . I set up benchmarks and started testing what works fast, I honestly got tired of hours of benchmarking and did what usually developers do right now when there are in a dead-end - asking LLM, I asked my agent to make it fast as possible, usually I don't do this, but this time and decided to establish proper tests and benchmarks, so AI can comfortably deprive me of my job. I let it iterate for some time, while going out for bread. Long story short - it did improve the performance, but it also told me that to push limits even further, it would need to sacrifice some specification and "correctness", which I thought is insane since I felt like it can be much faster. I tried some times again and again until I noticed that it's cycling 2-3 the same implementations and ignoring my not