Does prompt caching actually save you meaningful money on AI agents?

I’ve been thinking about prompt caching for agent workloads. In theory, it sounds very useful because agents often resend the same system prompt, tool schemas, instructions, memory, and context across multiple turns. If those repeated tokens can be cached, the cost savings should be meaningful. But I’m not sure how well this works in real production setups. A few questions I’m curious about: - Do you actually see meaningful cost savings from prompt caching? - Does it work well when agents use long tool definitions or large system prompts? - How much does multi-provider routing hurt cache hit rates? - Would you prefer sticky routing for better cache locality, or flexible routing for availability and latency? - At what scale does prompt caching become worth designing around? For people running real AI agents in production, is prompt caching a major cost lever, or is it more of a small optimization compared with model choice, context trimming, batching, and better task routing?