The post describes an agent request as a combination of tool definitions, developer instructions, conversation and tool-call history, and the newest turn.
It says caching saves the model's prefilling state, including key-value vectors, for an identical prompt prefix. A subsequent request can load that state and process only the newly appended content. Bedrock can choose a reusable prefix implicitly, while developers can mark one explicitly with prompt_cache_breakpoint; AWS recommends using both approaches together.
The examples focus on keeping stable instructions and tool definitions ahead of changing conversation material, including cases where tools change or an agent starts helper tasks. AWS presents the technique as a way to reduce repeated input processing, latency, and input cost.
The source does not provide an independent benchmark, a universal cost reduction figure, or evidence that the pattern transfers unchanged to APIs outside Amazon Bedrock. It is an AWS-authored implementation guide rather than a product-release announcement.