October 31, 2026

AWS published guidance on 26 October 2026 on how agents on Amazon Bedrock stop paying full price to resend their own history. An agent sends its tool list, instructions and every earlier result on each step. AWS reports that on GPT-6 Astra, implicit caching gave about 62% lower input cost than no caching in a six-turn workload, and explicit breakpoints about 61%.

Why it matters: Input tokens are the part of an agent bill that grows with every step. If your finance lead asks what a ticket costs now against last quarter, the cached share of input is the line to watch. AWS's own post tells readers to measure their agent, because these figures come from one workload and one model.

The catch: The saving is fragile. The shared start of the prompt must reach 1,024 tokens and match exactly, and AWS says adding, removing or reordering tools can invalidate it. The post reports input cost only, not total cost or speed, and the sections I read give no price for writing to the cache. AWS also warns that caching can expose tools and system prompts, and asks teams to cover it in their responsible AI practices.

What to do: Sort your tool definitions into one fixed order and move anything that changes, such as the date or user name, below the instructions. Add tools mid-session with the additional_tools input item instead of editing the list. Then log cached and uncached tokens on each request for a week and compare the result with your invoice before you promise anyone a saving.