Introducing Amazon Bedrock AgentCore Runtime Instances for Complex AI Workloads

According to the AWS News Blog, these instances allow agents to maintain state for up to 14 days, access GPU acceleration, and collaborate on a shared host, removing the need for developers to manually provision and manage the underlying networking and scaling infrastructure.

Managed Infrastructure for Stateful Agent Workflows

The new runtime instances complement existing AgentCore runtime microVMs. While microVMs provide a managed environment for invocations lasting up to 8 hours, runtime instances target workloads that require continuous operation over several days or direct access to the underlying operating system. According to AWS, this infrastructure is purpose-built for agents that must persist state across multi-step workflows that span long durations.

Developers can deploy multiple agents within a single runtime, each maintaining its own dependencies and artifact types. These agents can operate within shared sessions that persist for up to 14 days. To ensure long-term recall beyond a single session, AWS integrates runtime instances with AgentCore Memory and Amazon Elastic Block Store (Amazon EBS).

GPU Support and Framework Integration

Runtime instances support GPU-accelerated instance types for compute-intensive tasks. AWS stated that this removes the previous requirement for developers to manually configure EC2 instances, networking, and session management when GPU access or multi-agent coordination was necessary.

The service is designed to be framework-agnostic. According to the AWS News Blog, developers can use any model and any framework, specifically naming CrewAI, LangGraph, LlamaIndex, and Strands. Deployment requires minimal packaging, utilizing a zip file or container image and a @app.entrypoint decorator.

Multi-Agent Coordination and Orchestration

The architecture allows for a hybrid compute model where runtime microVMs and runtime instances work together via the same AgentCore APIs. In this configuration, a lightweight orchestrator agent on a microVM handles API calls and task routing, while specialized worker agents on runtime instances execute heavy tasks such as GUI automation, security scanning, or code compilation.

Agents on runtime instances can collaborate through a shared file system within a session. This allows agents to interact without direct API calls or message exchanges. For example, a code writer agent can save a file to a shared session directory, which a code reviewer agent can then read and analyze autonomously.

Deployment Specifications and Availability

Deployment involves creating a capacity provider to define the EC2 infrastructure, including the operating system and instance type. AWS supports Linux (ARM64 and x86_64) at launch. The service supports Python runtimes 3.11 through 3.14 with native code support, as well as container images.

Pricing for the service consists of standard EC2 pricing plus a management fee for AgentCore orchestration. To manage costs during idle periods, the service includes a session stop and restart feature.

Amazon Bedrock AgentCore runtime instances are available in the following regions:

• US East (N. Virginia, Ohio)
• US West (Oregon)
• Europe (Ireland, Frankfurt)
• Asia Pacific (Tokyo, Sydney, Singapore, Mumbai)