Agent Swarm Economics: Why Caching Was 78% of Our Bill A 24-hour measurement of an agent fleet running Claude Opus 5.5 at API list prices between 2026-11-15 and 2026-11-17 found that 78% of the Claude Code bill came from caching, because a loop that keeps appending to one session re-reads its entire growing context on every turn. The author's fixes include a context limit or a handoff skill that records current priorities so a fresh session can restart from them, and pruning the system prompt, which the author estimates would save about $93 a day in API spend and up to about $155 with a leaner orientation. The post argues long context windows suit one-shot questions rather than loops, and that swarm structure determines how many tokens go to messages, approvals and handoffs that do not move work forward. Agent Swarm Economics: Why Caching Was 78% of Our Bill November 17, 2026 Caching looks cheap, but in a long agent loop it quietly becomes most of your token bill. A million-token context window is mostly a trap. I ran our agents in a naive loop and measured what it cost. At API list prices, 78% of the Claude Code bill was caching