Guide from AWS combines implicit and explicit prompt caching for agents on Amazon Bedrock
AWS describes combining two prompt caching approaches for agents using OpenAI models on Amazon Bedrock. An explicit boundary allows shared instructions and tool definitions to be reused, while implicit mode continually handles the growing conversation.
AWS has published an implementation guide on combining implicit and explicit prompt caching for AI agents using OpenAI models on Amazon Bedrock. When the input begins with an identical sequence, the service reuses the cached KV state, meaning previously computed representations of input tokens. According to AWS, this can reduce the cost of repeatedly processing input and speed up agent responses.
In implicit mode, Amazon Bedrock automatically selects the end of the reusable portion. Explicit mode marks the boundary using prompt_cache_breakpoint and allows stable instructions and tool definitions to be shared across tasks, sessions, and subagents. The guide recommends combining both approaches: explicitly define the shared foundation and use implicit mode for the growing history. The reusable portion must contain at least 1 024 tokens, and its text must be exactly identical. Variable data should come after the marked boundary.
In one internal test by AWS, 4 120 tokens of instructions and tool definitions were cached on the first request, and subsequent requests loaded those 4 120 tokens from the cache. This is the result of a specific test, rather than evidence of a general savings rate. Details are available in the source article.
Why it matters
The guide addresses repeated processing of the same foundation during subsequent agent steps and when starting another task. Implicit caching of the entire conversation so far may not, on its own, preserve a separately reusable foundation of instructions for different tasks; an explicit boundary defines that foundation.
Two audiences, two different impacts
What this means
For individuals
The agent developer must preserve exactly the same wording and order of shared instructions and tool definitions to reuse their cached state.
For a business
According to AWS, this approach can reduce input costs and response times for enterprise agents with repeated instructions and shared tools. The cited test demonstrates repeated loading of tokens from the cache, rather than specific financial savings in enterprise use.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.