when you design REST, don't stop at resources. let your API guide the client through HATEOS the core of fielding's where "state transfer" means navigating between linked states not merely transferring data.
treat KV cache hit rate as a first-class production metric: in agentic systems your prompts are long and repetitive by nature. same system prompt, same tool schemas, same history prefix on every turn. a low KV cache hit rate means you're paying full prefill cost on every request. keep your prompt prefixes stable, structure consistent, and avoid unnecessary context churn. it's one of the cheapest wins you'll get when scaling agents.