The Pulse

AWS expands AgentCore runtime for larger AI agent sessions

AWS has raised AgentCore runtime session limits to up to 8 vCPUs and 32 GB and added a pricing option for always-on agents. The company says the updated runtime also reclaims released memory and lets agents mount filesystems.

AI.info Team · · fact-checked · against 3 sources ↓ · Published 12 Nov 2026

Amazon Bedrock AgentCore runtime sessions can now use up to 8 vCPUs and 32 GB of memory, alongside a new billing option for always-on agents. AWS announced the changes to its managed environment for deploying and running agents in a November 12 post.

The company says session startup is now up to 15 times faster and sessions can scale to four times more compute. The update also supports container images of up to 10 GB and lets agents mount filesystems to save their work.

Bigger sessions and faster starts

AgentCore runtime provides a managed compute environment in which developers can deploy and run agents without building or maintaining the underlying infrastructure. Each session runs in a hardware-isolated Firecracker micro virtual machine with its own identity, policy and audit trail.

The new generation of the runtime was introduced on September 18. AWS said it changed how sessions use memory and how agent environments load: instead of repeating initialization for every new instance, the service captures a snapshot of an initialized environment and restores it when an instance starts.

In its latest post, AWS says sessions start in under two seconds regardless of image size and concurrency. The company previously described a test using an empty echo agent that called no model or tools, with 5,000 cold invocations per agent across both runtime versions and five image sizes.

AWS said that test produced a P75 cold-start latency of about two seconds for the new runtime across images from 200 MB to 2 GB. The original runtime ranged from roughly 5.4 seconds to nearly 30 seconds; the client-side measurements included a round trip between AWS Regions over the public internet.

Memory and billing changes

The runtime reclaims memory within minutes after an agent releases it, and an idle agent scales to zero. For multi-turn agents with pauses between interactions, AWS says this can mean up to 68 percent less memory consumed per session.

A committed baseline pricing option sets a per-session memory baseline for always-on agents and bills that amount as a minimum. The session can burst on demand to up to four times the baseline, with pricing at AWS Lambda MicroVM rates.

Developers can select the new runtime by setting the platformVersion parameter to V2 when creating or updating a runtime. AWS says the next-generation runtime is available in regions across North America, South America, Europe and Asia Pacific.

Storage and customer results

Agents can also now mount and save work to a filesystem, including an Amazon Simple Storage Service bucket with FUSE support. AWS says the data remains within the session's trust boundary under customer-defined controls.

Per-session scoping remains planned: AWS says that by re:Invent, session-labeled credentials will bind each session to its own storage folder. That capability is separate from the filesystem support described in the update.

In the announcement, Coinbase VP of Engineering Siwei Shen said migration reduced the company's agent-session cold starts from nearly a minute to under five seconds, including during peak traffic. Shen also said startup delays longer than 30 seconds were eliminated while weekly agent volume more than doubled.

Mission Cloud reported results from an initial comparison of the same workload across both runtime versions. Chief AI & Data Scientist Dr. Ryan Ries said inference-request cold starts were 35 percent faster, health-check starts were 77 percent faster and measured runtime cost per call fell 17 percent.