AWS broadens regional access to its next-generation AI agent runtime

AWS is making its next-generation Amazon Bedrock AgentCore runtime available across regions in North America, South America, Europe and Asia Pacific. In its November 12 announcement, the company detailed larger computing sessions, faster starts and lower memory consumption for AI agents, while retaining a separate hardware-isolated environment for every session.

AgentCore runtime is AWS’s managed computing environment for deploying and running agents without maintaining the underlying infrastructure. This is not its first release: AWS announced the next-generation runtime on September 18, initially listing five regions. That earlier announcement already included elastic memory management, hardware-enforced isolation and consistent startup times. November’s announcement describes wider geographic availability alongside capacity limits, storage access and customer deployment results.

Sessions now support up to eight virtual CPUs, 32 GB of memory and container images up to 10 GB. AWS describes the computing capacity as four times the previous version. The company points to compiling large codebases, browser-driven workflows and running models locally as workloads that need the extra room.

AWS says sessions start in under two seconds, regardless of container image size or the number starting concurrently. Its September explanation describes the mechanism: prepare an agent environment once, save a snapshot, then restore new instances from that snapshot rather than repeat the full startup sequence.

The earlier AWS test measured cold starts—the delay before a fresh environment can begin work—at 1.9–2.0 seconds at the 75th percentile. That test covered images from 200 MB to 2 GB, versus 5.4–30 seconds on the previous runtime. Coinbase’s deployment account, included in the November announcement, offers a different workload-specific result rather than the same benchmark.

Migrating to the next generation of AgentCore runtime reduced our agent session cold starts from nearly a minute to under five seconds, even during peak traffic. It also eliminated startup delays longer than 30 seconds, all while our weekly agent volume more than doubled.

Siwei Shen, VP of Engineering, Coinbase, in AWS’s announcement

The billing changes address agents that pause while waiting for a model response, a tool or a human decision. The runtime reclaims memory within minutes after an agent releases it, and idle agents scale to zero. AWS says multi-turn agents with pauses between interactions can consume up to 68% less memory per session. That figure describes memory consumption, not a blanket reduction in the total bill.

For always-on agents, AWS offers committed baseline pricing. Customers set a memory baseline per session and pay for it as a minimum, with the ability to burst to four times that baseline. AWS positions this as a predictable rate for steady workloads, using AWS Lambda MicroVM rates. The commitment therefore carries a minimum charge rather than billing solely for each moment’s consumption.

Each session continues to run in its own Firecracker micro virtual machine, a small hardware-isolated computing environment, with its own identity, policy and audit trail. Agents can now dynamically attach a filesystem and save work to it, including an Amazon S3 bucket through FUSE support. AWS says this brings storage under the controls governing the job.

A more specific storage safeguard remains planned. AWS says that by re:Invent, session-labeled credentials will bind each session to its own storage folder. That distinction separates the filesystem access available now from the promised mechanism for restricting a session to its designated folder.