API connectivity is a way we might address /grounding/ to get up-to-the-moment information to incorporate into the LLM and to test whether a proposed action can execute successfully. As for world-models, I'm not sure it helps a lot. API access would give access to more real-time real world data. World models must be learned, so that requires a feedback loop with sensors and effectors. Accessing sensors through an API will be slow relative to having texts readily available in a giant, pre-prepared dataset. Accessing effectors (because world models map state and action to new state), is the bottleneck.