Give Your Sub-Agent Team Prompt Caching — From Re-Paying for Every System Prompt to a High Cache-Hit Rate in One Session

Paste this into Claude Code, Cursor, or Aider and it'll audit every place your agent team calls the Anthropic API, find where you're re-sending large system prompts and tool schemas at full price on every sub-agent call, and add cache_control in the right places — the shared tool loop, the main conversation history, and the per-agent one-shots — while catching the silent invalidators (a timestamp in the system prompt, an unstable tool order) that quietly keep your hit rate near zero. Audit first, then fix tier by tier. Each fix ships independently with before/after verification from the usage numbers you already get back.