NewArchitecture
15 min read1 views
Prompt Cache Invariants: The Discipline That Keeps an AI Daemon Cheap
Why one cache-busting mutation can 10x a conversation — and the rules that stop it
Every LLM provider keys its prompt cache on the prefix of the request. Mutate the system prompt, tool schema, or message history mid-conversation and you invalidate the cache for the rest of the session — a single slip can 10x the cost. The four forbidden mutations, the permitted ones, and how to enforce them statically.
ArchitectureAIPerformanceCost OptimizationAgents
