AWS announced an updated AgentCore runtime on November 12, increasing session capacity to 8 vCPUs and 32 GB. Its release post also describes memory reclamation, idle scaling and filesystem support for longer-running agents.

The change addresses a practical deployment problem: agents wait for models, tools and people, but their execution environments can keep consuming resources. AWS is trying to reduce that idle overhead while preserving separate execution environments.

These are AWS's product and performance claims. The announcement does not establish that every workload will receive its maximum startup or memory improvement.

Larger sessions retain the microVM boundary

AgentCore runtime is the managed compute layer within Amazon Bedrock AgentCore. AWS says each session continues to run in its own hardware-isolated Firecracker micro virtual machine, with separate identity, policy and audit records.

The new session limits are four times the previous version's compute capacity, reaching 8 vCPUs and 32 GB of memory. Container images can be as large as 10 GB. The target workloads include large repository compilations, browser workflows and local model inference.

AWS says sessions start in under two seconds regardless of image size or concurrency, and describes startup as up to 15 times faster. These statements describe the provider's reported behaviour; they are not measurements performed for this article.

The architectural trade-off is familiar. Pre-provisioned environments can reduce waiting but consume resources while idle. On-demand environments can save capacity while introducing startup delay. AWS positions the new runtime as a way to reduce that trade-off without customers maintaining warm pools or clusters themselves.

Memory savings depend on pauses and released resources

The runtime now reclaims memory within minutes after an agent releases it, according to AWS. An idle agent scales to zero. For multi-turn agents with natural pauses, AWS reports up to 68% less memory consumed per session.

That qualification matters. The 68% figure concerns memory consumption for workloads with pauses, not an unconditional reduction in total operating cost. Model calls, tools, storage and workload behaviour can affect the broader bill.

For always-on workloads, committed baseline pricing sets a minimum billed memory amount per session. Sessions can burst to four times that baseline. A minimum commitment and idle scaling serve different workload patterns; they should not be treated as the same economic arrangement.

The announcement also describes restoring initialized environments as part of the efficiency changes. Teams evaluating the service still need measurements using their own initialization work, active periods and pauses.

Customer reports use different baselines

Mission Cloud reports an initial same-workload comparison between runtime versions: 35% faster inference cold starts, 77% faster health-check starts and 17% lower measured runtime cost per call. These are customer-reported results included by AWS.

Zuora separately reports around 60% less memory per invocation in many use cases, while handling more traffic at steady latency. That is a different measure and workload. It cannot be combined with Mission Cloud's cost-per-call result into a universal savings figure.

The qualitative examples also cover distinct applications. Coinbase describes isolated coding-agent sessions. Rescale describes agents that validate simulation inputs, troubleshoot jobs, select hardware and generate reports, with engineers retaining control. Thomson Reuters describes using the runtime to host and scale CoCounsel legal workflows while preserving control of operation and integration.

These examples explain why managed isolation is useful. They do not independently certify every security or reliability claim in the release.

Storage enters the same control discussion

Long-running sessions accumulate files, credentials and tool outputs. Isolating compute alone does not determine how that data is stored or who can access it.

AWS says agents can dynamically mount filesystems and save their work, including to an Amazon S3 bucket using FUSE support. The release presents this as extending the trust boundary to data under a shared set of controls.

There is an important timing limit. The same announcement says per-session scoping, binding each session to its own storage folder through session-labelled credentials, is planned by re:Invent. That future commitment should not be reported as already available on November 12.

A session's isolation therefore still needs to be considered alongside configured identities and data-access policies. Hardware separation does not, by itself, establish that an application's permissions are correct.

Availability and the next verification point

AWS says the updated runtime is available in regions across North America, South America, Europe and Asia Pacific. The post does not provide a complete region-by-region list in its availability paragraph.

The immediate changes are larger isolated sessions, reported faster initialization and more flexible memory use. The next concrete storage milestone is the promised per-session folder scoping. For deployment teams, workload-specific latency, memory and permission checks remain the way to determine whether these changes meet their own requirements.

Related reading
Sources

Browse the directory →