Managing resources across IoT, edge, and cloud layers requires continuous, context‑aware decisions under constraints that rarely remain static. Deep reinforcement learning (DRL) is well‑suited for such problems, and large language models (LLMs) are increasingly employed to augment DRL pipelines, yet their architectural relationship is seldom made explicit. Building on Wang et al.'s taxonomy of Continuum Orchestration Systems that use DRL, we introduce two additional dimensions: the AI Augmentation Paradigm, which measures how LLMs are leveraged, and the Feedback Channel, which captures whether and through which system path execution feedback returns to the LLM to close the MAPE control loop at the LLM orchestration layer. Applying this taxonomy to six recent system architectures reveals a common gap: none combines full LLM orchestration with complete agent‑layer feedback in a Cloud Continuum setting. We attribute this gap to a missing cross‑tier feedback abstraction that would bridge the incommensurable per‑tier signals with the LLM orchestrator.
Review