You are selecting models for a multi-agent system: a lightweight classifier agent routes intents; a deep reasoning agent handles complex multi-step plans; a vision agent reads PDF screenshots. What architectural principle should guide model assignment?
Select an answer to reveal the explanation.
Short Explanation
You match each agent to the model family that fits its job: small/fast for routing, stronger reasoning for planning, and vision only when images matter. Don't throw one giant model at everything, force tiny models to do hard planning, or rotate models randomly.
Full Explanation
Architect skills include matching task demands to model family capacities. The correct principle right-sizes classifier, planner, and vision roles: small/fast models for routing, stronger reasoning models for planning, and vision-capable models only where image understanding is required. Assigning the single largest multimodal model to every agent ignores cost and latency. Using only on-device micro-models for all planning over-constrains capability. Randomly rotating models removes deliberate capacity matching. Include evaluation gates when swapping model families so routing and planning quality remain bounded.