How to use cache_control to cut Claude API input costs by up to 90% on agents with large stable prompts — the prefix-match invariant, what to cache, silent invalidators, and the break-even math.
Updated for 2026. The cost math I use to pick Claude Haiku vs Sonnet per task — token economics, the routing rule, and where cheap models quietly cost more.
Updated for 2026. The exact playbook I use to debug AI agents running live — trace every step, isolate the failing layer, and reproduce non-determinism fast.
Updated for 2026. The orchestration patterns I use to run multi-agent systems — durable queues, externalized state, and clean handoffs that survive retries.
Updated for 2026. The eval harness I use to ship AI agents — a graded test set, an LLM judge, and a regression gate so a prompt change can't break prod.