Stream processing has been and continues to be an essential compo- nent in the agent-centered information processing paradigm. Future agent archi- tectures are going to exploit large language models (LLMs) for fulfilling complex tasks in communicating and cooperating with other agents and humans. To over- come the well-known problems of LLM-based services (security, safety, trust- worthiness, explainability, human-beneficial behaviour), we propose equipping those architectures—among others—locally with a hierarchical model of under- standing. We discuss the consequences of this model for stream reasoning and illustrate them with an example of an information retrieval agent.