You are deploying a multi-agent customer-support system on Kubernetes using NVIDIA GPU nodes and Triton...

The rejected options are weaker because tuning one component in isolation or relying on FP32/default settings leaves GPU memory bandwidth, batching windows, and queuing delay unmanaged. Sub-100ms and zero downtime require GPU-aware autoscaling, latency metrics, health checks, and DCGM/Grafana visibility. CPU or memory-only scaling signals are too indirect. Option C is the correct engineering choice because the requirement is not just “make the model answer,” but control the execution surface. The selected option specifically C states “Deploy GPU pods in a node pool spanning all zones, mix GPU types, enable Cluster and Horizontal Pod Autoscalers using Prometheus GPU and latency metrics, and monitor with NVIDIA DCGM and Grafana.”, which matches the operational requirement rather than a superficial wording match. In NVIDIA terms, Triton’s metrics make GPU and model behavior visible enough to correlate batching efficiency with user-facing latency. That matters because measuring queue time, compute time, execution count, and memory pressure instead of guessing from average response time. The result is a system that can be benchmarked, traced, and revised without destabilizing the whole agent fabric.