Earendil’s prompt caching piece is important because it treats cache as architecture. Agent systems often carry a large stable prefix: instructions, tool schemas, policies, project conventions, and sometimes retrieved context. If that prefix can be reused cheaply and predictably, context stops being only a prompt-writing problem and becomes an infrastructure layer.
That changes the economics of agent work. Instead of paying the full context cost on every step, the harness can separate stable context from the task delta. The agent still receives the necessary frame, but repeated work no longer has to re-send the whole frame as if every step were new.
It also creates pressure for cleaner harness design. Caching rewards stable, well-factored context. A messy prompt that changes constantly is harder to cache and harder to reason about. A disciplined harness can make cost, latency, and behavior more predictable at the same time.
What becomes infrastructure?
The source title, “Prompt Caching in Agents”, names the useful unit: repeated agent work often depends on a stable prefix plus a changing task delta. Caching makes that split operational, because the harness has to decide which context is durable enough to reuse.
For New Runtime, this belongs with the broader context-efficiency story: the future of agents is not bigger prompts everywhere. It is deciding which context is persistent infrastructure and which context belongs only to the current turn.