Next-gen Bedrock AgentCore runtime goes default: sessions start up to 15x faster, billed memory down 68%
AWS has switched every new Amazon Bedrock AgentCore runtime to its next-generation architecture. Swami Sivasubramanian, AWS VP of Agentic AI, summed it up: agent sessions 'can now start up to 15x faster, scale to 4x more compute, and run up to 68% more efficiently', without giving up isolation. The concrete figures: startup for a 2 GB container image dropped from 29.7 seconds to under 2 seconds at P75. Sessions can use up to 8 vCPUs and 32 GB RAM, and container images up to 10 GB are supported.
The key mechanism is snapshotting. An AWS Machine Learning blog post published November 13 explains that runtime V2 snapshots an initialized environment once and restores it for each new session, so cold-start time no longer grows with image size. Each session still runs in its own Firecracker microVM, the lightweight virtualization AWS uses for Lambda, so agents running model-generated code are separated by a hardware boundary rather than a shared kernel. The post covers how to measure startup and offers four patterns for keeping expensive work off the first request. The memory savings come from more efficient handling of idle time in long, multi-turn sessions.
This is AWS's answer to a practical agent problem: agents run longer and do heavier work, but users won't wait 30 seconds for a session. It fits a broader AWS push this week that included Nitro Sandboxes for EKS, ECS and EC2, an AWS Emulator for local agent testing, and guides on per-user Bedrock cost attribution and NeMo Switchyard model routing. Google, Microsoft and specialist sandbox vendors are all competing for the same agent-hosting workloads.
The reaction was muted compared with the week's big stories; AWS's own launch posts drew modest engagement on X. The real test is whether the P75 numbers hold for customers with large, irregular images. Snapshot-restore also brings its own pitfalls around state that must not be shared between sessions, such as random seeds and credentials, which the architecture guidance tries to address.
