Artificial intelligence news.
Straight from the original announcements.

AWS details four ways to speed up Amazon Bedrock AgentCore sessions

November 13, 2026, 23:02By the AI News Bites Editorial TeamSource: AWS
AWS details four ways to speed up Amazon Bedrock AgentCore sessions
Image: AWS

AWS has outlined four architecture patterns for reducing session startup delays in Amazon Bedrock AgentCore runtime, including snapshot-based initialization, build hooks and pre-warming when users engage. The guidance is aimed at agents that need to respond quickly in interactive chat and voice experiences.

Key points
  • AgentCore runtime V2 snapshots initialized environments for reuse.
  • New sessions restore memory, while image files are loaded lazily.
  • Build hooks can prime large files, browsers and CLI resources.
  • Background pre-warming targets sub-second interactive startup.
  • AWS recommends comparing Latency with OverheadLatency metrics.

AgentCore V2 prepares the environment once

Amazon Bedrock AgentCore runtime provides each session with an isolated microVM and scales to zero when there is no work. AWS says users pay for the resources their sessions consume. With platformVersion set to “V2”, the runtime initializes a container once, creates a snapshot and restores that snapshot for each new session.

During a create or update operation, AgentCore starts the container in a fresh microVM and waits for a healthy GET /ping response, which must arrive within 120 seconds. It then clears the operating system’s file cache, freezes the microVM and stores its memory and disk as a snapshot. AWS says each create or update takes a few minutes while that snapshot is built.

When a session begins, the snapshot is restored into a new microVM. Imports, objects and clients held in memory at the time of the snapshot are available immediately. Files from the container image are loaded lazily, so the first read of a block comes from snapshot storage and is then cached closer to the host.

Four patterns target different startup needs

AWS recommends switching to V2 and completing initialization before app.run() when a session starting in about 2 seconds is acceptable. The company also advises keeping static work out of the first request. This includes importing libraries, creating clients and loading static data from the image.

Some agents cannot keep required resources in their own process. A browser binary or a command-line tool that runs as a separate process and exits may be read during initialization but be unavailable after the snapshot is restored. For these cases, AWS says developers can use a build hook to load and prime the resource while the snapshot is built. AgentCore runtime now supports build hooks.

The third pattern is restructuring the agent so work that belongs in initialization does not run inside the request handler. The fourth is pre-warming: when users engage with an application, such as by opening the app or a chat window, the application can start a session in the background. AWS presents this option for experiences that need sub-second startup.

AWS recommends measuring the source of latency

AWS separates the time needed to make a session ready from the time spent in the agent’s code, including model calls, tool calls, sub-agents and response streaming. AgentCore runtime publishes Latency and OverheadLatency metrics in the AWS/Bedrock-AgentCore namespace. Latency measures time to first byte, while OverheadLatency covers platform work such as starting the microVM and restoring the snapshot.

Developers can estimate their code’s time by subtracting OverheadLatency from Latency. Initialization is different on V2: it runs while the snapshot is built rather than during each request, so it does not appear in request latency metrics. AWS recommends logging initialization time before app.run() and reporting percentiles such as P90 and P99, or trimmed means such as TM90, instead of relying on a plain average.

For users, the guidance means faster startup depends on the agent’s workload. V2 can move repeated static setup into a reusable snapshot, build hooks can handle resources that do not remain in process memory, and background startup can address interactive cases where even a short delay is noticeable.

Related news

AWS upgrades AgentCore runtime with faster starts and stronger isolation
Products and toolsNovember 12, 2026, 18:59

AWS upgrades AgentCore runtime with faster starts and stronger isolation

AWS has announced updates to the Amazon Bedrock AgentCore runtime, giving AI agent sessions faster startup, more compute capacity and expanded isolation for data. The…

AWS launches direct Web Search API for Bedrock AgentCore
Products and toolsNovember 12, 2026, 00:13

AWS launches direct Web Search API for Bedrock AgentCore

AWS has introduced the Web Search API on Amazon Bedrock AgentCore, allowing applications to retrieve ranked web results through a single API call. The service is…

AWS outlines push-based updates for its Agent Registry
For developersNovember 11, 2026, 21:52

AWS outlines push-based updates for its Agent Registry

AWS has outlined an event-driven architecture that automatically synchronizes AWS Agent Registry records when MCP servers gain tools or change their schemas. The design…