Production agents fail financially (unsustainable cost) and functionally (wrong tool usage). A 5% accuracy gain at 20× cost isn't sustainable. Cost-quality tradeoffs measure performance against compute cost. KAMI framework (Nov 2025) provides composite metric: quality / normalized_cost. Pareto frontier analysis plots quality vs cost for multiple models, identifying where no alternative is both cheaper and better. Don't Break the Cache (Jan 2026) shows prompt caching reduces multi-turn costs by 90%. Agent re-reading 20K token system prompt every turn wastes 80% of tokens. Caching achieves 10-turn cost reduction from $4.80 to $0.52 (89% savings). Implement per-invocation cost tracking: extract token counts, multiply by pricing, log to CloudWatch, create alarms when average exceeds thresholds. Tool correctness evaluation validates tool selection and parameters. CCTU framework (March 2026) proposes hierarchical validation: deterministic checks catch 83% of errors (free, instant), extractors catch 12% (fast), LLM semantic validation catches remaining 5% (expensive, 200-400ms). Three levels: tool selection (deterministic, free), constraint validation (business rules, dates valid, 1-10ms), semantic correctness (LLM-judge, $0.001-0.003). Cascading pattern achieves full coverage at 20× lower cost than validating everything with LLMs ($0.12 vs $2.40 for 1,200 cases). KAMI benchmark shows Haiku is Pareto-optimal for most workloads (quality 0.82, cost $0.002/query) - 7× cheaper than Sonnet. Walk away with cost tracking, Pareto analysis, cascading validation, and CloudWatch observability.