The Pulse
AWS makes snapshot-based AgentCore V2 the default for new runtimes in select regions
New Amazon Bedrock AgentCore runtimes use a snapshot-based version by default in selected AWS regions. AgentCore Runtime V2 restores initialized environments for new sessions, and AWS reports faster startup for sessions using a 2 GB container image.

New Amazon Bedrock AgentCore runtimes now use a snapshot-based version by default in selected AWS regions, replacing repeated session initialization with a restored environment. In a November 13 technical post, AWS described how AgentCore Runtime V2 prepares an agent once and restores that state for each new session.
AWS says sessions using a 2 GB container image start in under 2 seconds at the 75th percentile, compared with 29.7 seconds on the previous generation. The company made V2 the default for new runtimes in the listed regions on November 12.
Where the default applies
Existing runtimes remain on their current version until customers move them. Customers can select V2 when creating or updating a runtime, including in a region where it is not yet the default; they can also select V1 for a new runtime.
The default applies in regions including US East, US West, Canada Central, South America and several European and Asia Pacific locations. AWS also lists locations such as Europe Milan, Spain and Stockholm, and Asia Pacific Malaysia, Thailand and Tokyo.
How the snapshot is made
Under V1, AgentCore pulls the container image, starts it and runs initialization code before a new session can process its first request. With V2, the runtime starts the container when the runtime is created or updated and waits for a healthy response to GET /ping, which must arrive within 120 seconds.
It then clears the operating system file cache, freezes the microVM and stores its memory and disk as a snapshot. Creating or updating a V2 runtime takes a few minutes while that snapshot is built.
Each new session restores the snapshot into a fresh microVM, with the process resuming alongside its initialized objects and clients. Image files are loaded as needed: a block is fetched when the session first reads it. The file cache does not carry over, so reading a file during initialization without retaining its contents does not make that read fast after restoration.
Changes for agent builders
The post recommends moving reusable work such as imports, client creation and loading static data into initialization rather than leaving it for the first request. Data kept in a process object or on a RAM-backed file system is available after the snapshot is restored.
Network connections require different handling: open connections become stale after restoration. AgentCore also supports build hooks for cases such as browser binaries or command-line tools that cannot hold what they need in the agent process. Holding more data in memory makes the snapshot larger and incurs charges for that memory.
Measuring the startup
AgentCore publishes Amazon CloudWatch metrics for time to first byte and platform overhead. Initialization time is not part of a V2 request: it occurs during snapshot creation and adds to the time a runtime takes to become ready.
The runtime also scales sessions to 8 vCPU and 32 GB of memory and supports container images up to 10 GB. AWS says memory is reclaimed within minutes of release and that multi-turn agents can see up to 68% less billed memory per session.
Sources
Companies in this story
Related stories
- AWS expands AgentCore runtime for larger AI agent sessions
- AWS details per-user cost tracking and limits for Bedrock AI tools
- Amazon Bedrock adds search across spreadsheet tables
- H Company makes AI agents available on AWS European Sovereign Cloud
- AWS details OpenFold3 research preview deployment on SageMaker