All an LLM does is predict the next token based on input tokens.
Most times in response to a user request, the Agent forwards the request to the LLM, then the LLM goes to the Agent and says “I don’t have that information, I need you to get it for me”, and the Agent basically does a web search and sends the information back to the LLM. You can engineer it, to direct from where the Agent will pull the information it sends back to the LLM.
Claude does this flawlessly. Anthropic > OpenAI in the ugliest way. And what does Flim Flam do? Target ads coming to your Codex terminal.