A development sequencing rule: prototype and validate a new agent capability using the largest/most capable model available, and build evals and success criteria against that ceiling of performance. Once evals confirm the behavior meets the bar, swap in smaller, cheaper models and re-run the same evals to see how far down the model size can go without regressing. This avoids two failure modes — abandoning a capability because a small model couldn't do it, and prematurely optimizing cost before knowing what 'good' looks like. Pairs with Eval-Building Workflow (Dataset → Prompt → MCP → Scorer) and Model Tier Selection by Task Upside.