When evaluating a multi-agent customer service system experiencing unpredictable scaling costs and performance bottlenecks during...

For this scenario, the combination of Options D and E is defensible because it exposes the control plane that a senior engineer can test, scale, and harden. Cost attribution and workload profiling show which agent type consumes GPU time and whether batch sizing or HPA thresholds are wrong. Constant allocation hides waste. Operationally, the design depends on profiling the request path from ingress through guardrails, routing, Triton scheduling, TensorRT-LLM execution, and response assembly. Together, D states “Deploy distributed tracing with cost attribution per agent type, correlating resource consumption with business value metrics to identify optimization opportunities in agent deployment strategies.”; E states “Implement comprehensive workload profiling using NVIDIA Nsight to analyze GPU utilization patterns, identify underutilized resources, and optimize batch sizing for dynamic scaling with Kubernetes HPA.”, so the answer covers both sides of the requirement instead of solving only the model or only the infrastructure layer. The alternatives would look simpler in a prototype, but overlarge batches may improve throughput while violating interactive latency targets. Within the NVIDIA stack, NVIDIA Perf Analyzer, GenAI-Perf, Nsight, and Triton metrics help isolate whether the bottleneck is batching, compute, memory, or request scheduling. It also creates clean evidence for audits, incident review, and root-cause analysis when behavior drifts.