AWS has launched the next generation of Amazon Bedrock AgentCore Runtime, its serverless microVM environment for deploying production AI agents.
Runtime V2 introduces elastic memory management: memory is allocated as needed and reclaimed when it becomes inactive, instead of billing a session against its peak footprint until it ends. AWS says this should reduce costs for long-running or bursty agents.
AWS also replaced repeated container initialization with snapshot-based startup. In its benchmark, V2 delivered 1.9–2.0-second P75 cold starts across container images from 200 MB to 2 GB. V1 ranged from 5.4 to 30 seconds.
The test used an empty echo agent that called no model or tools, so these figures measure infrastructure startup—not complete agent response time. AWS ran 5,000 cold invocations per agent across both runtime versions and five image sizes.
Developers can adopt the runtime by setting platformVersion to V2. It is currently available in five AWS regions: Northern Virginia, Ohio, Oregon, Ireland, and Tokyo.
Why it matters
Models and orchestration frameworks receive most of the attention, but production agents also need responsive, isolated, economical compute—especially when they run unattended for hours or remain idle between events. AgentCore Runtime V2 directly addresses that infrastructure layer, making scale-to-zero agent deployments more practical without maintaining pools of warm containers.
The release also shows how managed-agent competition is moving below the interface layer. Google’s updated Antigravity managed agent packages tools and a hosted sandbox; AWS is optimizing the microVM substrate that keeps similar workloads isolated, responsive, and economical at scale.