Wide World Importers is selecting models for a multi-agent supply-chain suite in Microsoft Foundry. Tasks include: (1) high-volume SKU classification from short structured fields, (2) multi-step negotiation email drafting with careful tone, and (3) overnight batch reconciliation of millions of line items with strict cost caps. Which model capacity matching approach is most appropriate?
Select an answer to reveal the explanation.
Short Explanation
Match the model to the job. Short structured classification wants small/fast; careful multi-step drafting wants a stronger reasoning/writing model; huge overnight reconciliation wants cost-efficient throughput under caps. That is option B. One frontier model for everything (A) burns budget on easy tasks. A VLM for text-only work (C) is the wrong family. Regex-only (D) abandons agentic capability where language flexibility is required.
Full Explanation
Architects must match task demands to model family capacities. Option B correctly differentiates latency/cost-sensitive classification, quality-sensitive writing/reasoning, and batch cost-constrained reconciliation.
Option A ignores cost and latency budgets; frontier models are rarely optimal for high-volume trivial classification.
Option C mismatches modality and wastes capacity; vision-language models are for multimodal inputs, not a default for text tabular classification.
Option D underfits complex negotiation language and exception-heavy reconciliation narratives that benefit from LLM agents with tools.
Exam tip: Multi-agent systems often use heterogeneous models per agent role rather than one global model.