Amazon has released a new platform version of the Bedrock AgentCore runtime, the managed service on which agents built with its tooling run. Two things change. Memory is now reclaimed as a session releases it, rather than the bill tracking peak allocation for the session’s entire life; AWS says the rate per gigabyte-hour is higher but the number of gigabyte-hours is far lower, which should reduce cost for most agents. Separately, containers now start from a snapshot, so cold-start time no longer scales with image size.
Pricing mechanics are not the most exciting subject, but for long-running agents that sit idle between tool calls they matter more than model choice. An agent that holds a session open for an hour while waiting on a human has been paying for its peak footprint throughout.
The latency figures are AWS’s own and describe a deliberately trivial case: an echo agent making no model calls, measured over 5,000 cold invocations, at roughly two seconds at the 75th percentile for images from 200MB to 2GB, against 5.4 seconds to nearly 30 seconds on the previous runtime. Real agents will be dominated by other costs. It is opt-in, by setting the platform version to V2. Details are on the AWS machine learning blog.