Amazon Web Services said the next generation of its AgentCore runtime became the default for new deployments on November 12, 2026, replacing a predecessor that the company says took 29.7 seconds to start a session on a 2 GB container image.
The new runtime, announced in an AWS What’s New post1, restores sessions from a pre-initialized snapshot rather than starting them cold. AWS claims startup time drops to under 2 seconds at the 75th percentile for the same 2 GB image. The date is drawn from the timestamp on a companion AWS Machine Learning Blog post2 about the runtime; the blog’s body text could not be retrieved, so product details rest on the What’s New announcement.
AgentCore, the managed service that the runtime underpins, is part of Amazon Bedrock. It deploys AI agents built with frameworks such as CrewAI, LangGraph, LlamaIndex, Google ADK and the OpenAI Agents SDK3 inside Firecracker microVMs that provide hardware-level isolation. Sessions scale to 8 vCPU and 32 GB of memory and support container images as large as 10 GB.
Idle scaling and billed memory
Once a session idles, the new runtime scales it to zero and reclaims memory within minutes. For multi-turn agent workloads, AWS stated, that architecture reduces billed memory per session by up to 68 percent.
Existing runtimes remain on the current generation until customers migrate them. Administrators who want to adopt the new runtime in a Region where it is not yet the default, or to upgrade an existing deployment, set platformVersion to V2 when creating or updating a runtime. To keep a new runtime on the earlier generation, they set platformVersion to V1.
Rollout and what’s missing
The next-generation runtime is the default across 19 AWS Regions, including the US, Canada, South America, Europe and Asia Pacific. The announcement lists those Regions but does not state whether AWS plans to extend the default to additional ones.
The What’s New post does not include independent benchmark results, details on snapshot consistency guarantees, or specifics about how the memory reclamation mechanism behaves with long-running, stateful agent sessions. AWS directed users to its pricing page for cost details.