The whole topic is about training time, not the tokens you feed it during runtime. Don't think that it can reason about the stuff it was trained on the same way it can reason about tokens you give it. They can't. Any arguments based on what they can do with runtime token input are therefore not relevant.

Your objection is getting less and less clear. The issue is the LLM's capabilities at runtime given that it was trained to predict tokens. LLM agents demonstrate that such a training regime induces a system that can capably follow instructions. Nothing you've said is counter to this.