The implementation detail that matters is measuring queue time, compute time, execution count, and memory pressure instead of guessing from average response time. This is a lifecycle problem, not a wording problem, and Option B gives the team a controllable lifecycle for the agent behavior. Multimodal latency is a pipeline property. Profiling text, image, and voice paths together reveals switching overhead, queuing, and dynamic batching opportunities. For a production build, Triton's metrics make GPU and model behavior visible enough to correlate batching efficiency with user-facing latency. The selected option specifically B states
"Profile end-to-end latency across modalities, measure model switching overhead, analyze batch processing opportunities, and evaluate Triton's dynamic batching for multi-modal workloads.", which matches the operational requirement rather than a superficial wording match. 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. That is the difference between an agent that works in a notebook and an agent that remains reliable in production.