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.