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 demonstrat

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.