Caching in LLM agent systems

I read Luke Parker’s post and ended up looking more closely at a question I keep bumping into: if an orchestrator is long-running, does it keep getting prompt-cache benefits? The distinction that helped me is that the thread, the harness context, and the provider cache are separate things. A dormant thread is not spending tokens or keeping a provider cache warm. When it wakes, the next request can still hit if the early prefix is unchanged and the provider’s current retention and routing rules line up, but compaction, changed tools, new early instructions, or elapsed time can make it cold. I wanted one guide that keeps the model’s in-generation KV cache, cross-request prefix caching, semantic caching, and ordinary tool caches from blurring together.