Unlike the OP, this makes using an agent actually sound useful.

My summary of the article: I use a really cheap LLM, and suffer like everyone else using GPT 3.1. This is definitely a throwback to the days not so long ago when LLMs were completely incompetent.