Skip to content

All articles

AWS makes faster AgentCore the default for Bedrock AI agents

AWS has upgraded its AgentCore runtime for Amazon Bedrock, dropping start times to under two seconds and cutting billed memory by up to 68%.

On November 12, 2026, AWS announced (opens in a new tab) that its next-generation AgentCore runtime is now the default execution environment for autonomous AI agents built in Amazon Bedrock. The update addresses two of the most persistent barriers to deploying AI agents in production environments: latency and infrastructure costs. By changing how agent sessions are initialized and managed, AWS has reduced the P75 start time for a 2 GB container image from 29.7 seconds to under two seconds. For small and mid-sized businesses building customer-facing applications or internal automation tools, this shift turns complex AI workflows from expensive experiments into viable operational software.

The architecture of the new runtime

The technical foundation of this update relies on how AWS provisions compute resources for AI agents. AWS reported that each autonomous AI agent session in the next-generation AgentCore runtime executes inside a hardware-isolated Firecracker microVM. This isolation ensures that data processed by one agent cannot leak into another session. Security remains a primary concern for businesses deploying AI, and hardware-level isolation provides a necessary boundary when agents handle sensitive customer data or proprietary codebases.

Beyond security, the architecture changes how sessions begin. Instead of starting from scratch, the updated environment boots sessions from a saved snapshot of the initialized agent. This approach is what allows the P75 session start times for a 2 GB container image to drop so significantly. In software development, a 29.7-second delay is often unacceptable for a user waiting for a response. A delay of under two seconds is entirely manageable for conversational interfaces or background automation tasks.

The execution environment also supports significantly heavier workloads than earlier iterations. AWS announced that AgentCore agent sessions scale up to 8 vCPUs and 32 GB of memory while supporting container images up to 10 GB. This capacity allows developers to run demanding tasks directly inside the agent session. An agent can now compile a large codebase, drive a web browser for automated testing, or load a local machine learning model without requiring external compute clusters.

What it means for your business

For a company of 5 to 500 people, cloud infrastructure costs require careful management. The most immediate impact of the new AgentCore runtime is a reduction in the idle time businesses pay for. AWS reported that idle AgentCore agent sessions automatically scale to zero. This mechanism delivers up to a 68% reduction in billed memory per session for long-running, multi-turn conversational agents.

When an AI agent engages in a multi-turn conversation, there are natural pauses while the human user reads a response and types a reply. In previous architectures, the compute resources required to maintain the agent's context remained active and billed during these pauses. Scaling to zero means the meter stops running when the user stops typing. Memory is reclaimed within minutes of release. For a business processing thousands of customer service interactions daily, a 68% reduction in memory billing directly improves the profit margin of the application.

The increased capacity also simplifies how a small technology team builds software. Previously, running a heavy task like a browser automation script required the AI agent to call out to a separate server or serverless function. Managing these external resources added complexity, maintenance overhead, and additional points of failure. By supporting large container images and providing substantial compute power directly within the isolated microVM, AWS allows developers to consolidate their architecture. A smaller team can build more capable agents without managing a sprawling web of supporting cloud services.

This consolidation is particularly relevant for businesses in Asia building localized AI tools. A company in Singapore or Mumbai can deploy a heavy container image containing specific regional language models or localized compliance checking tools directly into the agent session. The infrastructure scales automatically with the work, removing the need for a dedicated operations engineer to monitor and adjust server capacity.

What to watch

The transition to the new runtime is automatic for new deployments, but existing systems require deliberate action. Older deployments remain on their original version unless a developer actively migrates them. To use the next generation, teams must set the platform version to V2 when creating or updating a runtime. Businesses with active Bedrock agents should audit their current deployments and plan a migration to V2 to capture the cost savings and performance improvements.

Availability is another factor to monitor. The new AgentCore runtime is now the default in several AWS regions across the globe. For businesses operating in the Asia Pacific region, the default status applies to Malaysia, Mumbai, Seoul, Singapore, Sydney, Thailand, and Tokyo. Companies operating in regions where the new runtime is not yet the default must manually specify the V2 platform version to access these features.

Teams should also watch how their container sizes affect the new start times. While AWS notes the start time of under two seconds for a 2 GB image, maximizing the 10 GB container limit will likely alter that performance metric. Technology leads should test their specific workloads to understand the exact latency users will experience before pushing a maximum-capacity container into a live production environment.

The maturation of AI infrastructure is moving away from raw model capabilities and toward operational efficiency. By addressing the fundamental mechanics of how agents start, run, and scale, AWS has provided a clearer path for businesses to deploy autonomous software. Faster start times improve the user experience, while automatic scaling protects the technology budget. Small and mid-sized companies can now build complex, heavy-duty AI agents with the confidence that their cloud bills will align closely with actual usage.

Sources

  1. AWS (opens in a new tab)aws.amazon.com

This article was written with AI and checked, sentence by sentence, against the sources linked above.