The selected option specifically C states “Deploy specialized NVIDIA NIM microservices with an LLM router to dynamically route requests to appropriate models based on complexity, combined with auto-scaling infrastructure that scales different model types independently.”, which matches the operational requirement rather than a superficial wording match. The decisive point is failure isolation: Option C keeps the agent’s decision path observable instead of burying behavior inside one prompt or one service. The runtime should therefore be built around independent scaling of agent components so embeddings, reranking, reasoning, and guardrails do not share one rigid capacity pool. Routing simple FAQs to cheaper models and complex reasoning to stronger models is the cost/performance sweet spot. Independent scaling avoids overprovisioning every agent tier. That is why the other options are traps: CPU-only or memory-only scaling signals rarely capture the saturation profile of GPU-backed LLM inference. The stack-level anchor is clear: NIM microservices and the NIM Operator fit Kubernetes production operations; Triton provides serving primitives and Prometheus-exportable inference metrics for GPUs and models. The answer is therefore about engineered control planes, not simply model capability.