Eric Schmidt recently commented that he thought three things would bring profound change: (a) an infinite context window, (b) chain of thought reasoning in agents and (c) the text-to-action capacity for programming.
In that view, agentic is mostly about (b) and (c). Having a way to chain together steps (which is just looping with the right context) and the ability to write and run code allows LLMs to take action on the things it is given to "do".
What you describe is basically just a standard LLM workflow. It's not as if the LLM is evolving the workflow itself as it goes.