Amazon Bedrock AgentCore V2 uses microVM snapshots to cut session startup times

Amazon Bedrock AgentCore runtime has updated its platform to V2 to speed up session launch times for artificial intelligence agents. The fully managed, serverless environment runs each agent session inside an isolated microVM. With the V2 update, the platform changes how it prepares these microVMs to reduce startup latency.

Snapshot technology replaces container initialization

In the older V1 version, AgentCore pulled container images from Amazon Elastic Container Registry and ran initialization code every time a new session started. This process caused cold starts, where large container image pulls slowed down initial request times.

Version 2 completes initialization once when a runtime is created or updated. The system runs the container inside a microVM, waits for a healthy status ping, clears the operating system file cache, and freezes the microVM into a snapshot. New sessions restore this snapshot directly into a fresh microVM. Memory contents, imported libraries, and pre-built software clients resume instantly from memory.

Optimizing code and using build hooks

Restoring a snapshot allows sessions to start in about two seconds for most setups. To get this speed, initialization tasks must run before the application starts listening for requests. However, file reads from disk images are loaded lazily when requested, and single-use command-line tools or browser binaries are not preserved in memory by default.

AgentCore runtime now supports build hooks to address this limitation. Build hooks allow developers to load and prime large files, external command-line tools, or web browsers while the snapshot is being built, keeping those assets ready in memory before sessions start.

Pre-warming sessions for real-time applications

For chat and voice applications, two seconds can still be too slow for end users. Developers can monitor startup performance using Amazon CloudWatch metrics, which track overall latency and runtime overhead.

To achieve sub-second response times, applications can use a pre-warming pattern. This approach sends a request to start the session in the background as soon as a user engages, such as when opening an application window, before they send their first message.