You are both right. An LLM inherently has access to stuff the same way a brain in a jar has access to stuff. It's information comes from fine-tuning the models to return syntax that agent code can interpret as a request to invoke a tool. That tool returns information to the context of the conversation. It doesn't learn and it can't truly remember things. Every time you start a session it is brand new. It sees your codebase for the first time every time.

The information access they have is whatever the agent allows it to access via tool exposure. Be it built in tools, or MCP servers

As I've mentioned elsewhere, not if by "information" you mean semantic content that a mind can process. What they have are vector fields (essentially just numbers) with statistically more or less likely relationships.

If I say, "take me out to the ballgame" to an LLM, the tokens representing the words in the next verse of the song are statistically "close" in the vector database, so it's likely to generate them. But that doesn't mean it actually knows the lyrics... or even has those lyrics recorded in a regular database anywhere.

That's why they hallucinate. The model determines that the next token is something nonsensical, but it has no way of understanding that it has made a mistake. In a sense, it actually hasn't made a mistake. It's done exactly what it's designed to do. It's just that in the case of hallucinations, its output isn't useful.