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 company says the managed runtime is designed to help organizations run agents securely and efficiently at scale.
- AgentCore sessions start in under 2 seconds, according to AWS.
- Sessions support up to 8 vCPUs, 32 GB and 10 GB container images.
- Idle agents scale to zero and memory can be reclaimed within minutes.
- Each session runs in a hardware-isolated Firecracker MicroVM.
- Agents can mount and save data through Amazon S3 with FUSE support.
Faster starts and larger agent workloads
AgentCore runtime is the managed compute layer of Amazon Bedrock AgentCore. AWS says it allows developers to deploy and run agents without building or maintaining the underlying infrastructure.
With the update, sessions start in under 2 seconds, independent of image size and concurrency, according to AWS. Sessions can scale to up to 8 vCPUs and 32 GB, compared with the previous version's capacity, and support container images up to 10 GB. AWS says this represents 4x more compute than the previous version.
The company also says sessions start up to 15x faster and run up to 68% more efficiently, while retaining hardware isolation. These capabilities are intended for workloads including large codebase compilations, browser-driven workflows and local model inference.
AWS says developers no longer need warm pools and pre-provisioned clusters to manage cold starts. Coinbase, Mission Cloud, Zuora and Rescale are among the organizations cited by AWS as using or evaluating the updated runtime.
Billing designed around actual agent activity
AWS is also changing how the runtime handles resources when agents are waiting for a model response, a tool call or a human decision. The runtime can reclaim 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 reduce memory consumed per session by up to 68 percent. For always-on agents, committed baseline pricing sets a memory baseline per session, with the ability to burst up to four times that baseline when needed. AWS says this provides a predictable rate at AWS Lambda MicroVM rates for steady workloads.
The company presents the changes as a way to align consumption with the work an agent performs rather than the time it spends waiting. AWS customer Zuora said the next-generation runtime cut its memory footprint per invocation by around 60% in many use cases while handling more traffic at steady latency.
The isolation boundary now includes data
Each AgentCore session continues to run in its own hardware-isolated Firecracker micro virtual machine, or MicroVM, with its own identity, policy and audit trail, according to AWS. The update extends that model beyond compute.
Agents can dynamically mount and save work to a filesystem, including an Amazon Simple Storage Service, or Amazon S3, bucket with FUSE support. AWS says this allows files, credentials, tool outputs and other data to remain inside the trust boundary for the duration of a job, governed by a single set of controls.
This addresses a challenge for long-running agents, which can accumulate context and require access to data outside their execution environment. AWS argues that isolation limited to compute leaves organizations to create their own rules for moving data in and out, increasing complexity and risk.
Why the update matters for production deployments
AWS is positioning the runtime for organizations moving from a small number of agents to thousands of concurrent sessions. The company cites research from METR saying that the length of tasks frontier agents can complete independently doubles about every four months. It also cites McKinsey, which found that 40 percent of large enterprises are scaling AI agents, up from 27 percent a year ago.
At the same time, AWS cites Deloitte research showing that only 15 percent of organizations have scaled orchestrated, multi-agent systems, while 70 percent say they cannot yet trust and govern their agents.
The updated runtime targets the infrastructure concerns behind those figures: isolated execution, faster startup, elastic compute, data controls and resource usage that can scale down during periods of inactivity. AWS's announcement suggests these capabilities are intended to reduce the operational barriers companies face when moving agentic workloads into production.


