AWS launches persistent compute for production AI agents

Amazon Web Services has rolled out runtime instances for Amazon Bedrock AgentCore, a new compute option designed to support production-grade AI agents requiring persistent infrastructure. The service targets developers moving beyond prototyping to deploy agents that maintain state across extended workflows, coordinate with other agents, or leverage GPU acceleration for compute-intensive tasks.

What runtime instances provide

Runtime instances offer AWS-managed EC2 infrastructure where multiple AI agents can operate within a shared environment. Each agent retains its own dependencies and artifact types while collaborating through shared sessions that persist for up to 14 days. Key capabilities include GPU acceleration for demanding workloads, session stop/restart functionality to reduce costs during idle periods, and support for containerized deployments. The service integrates with Amazon Elastic Block Store and AgentCore Memory, enabling agents to retain knowledge across sessions and environments.

Previously, developers needing persistent agent infrastructure had to provision and manage EC2 instances manually, including networking, session management, scaling, and monitoring. Runtime instances automate these tasks while maintaining compatibility with existing AgentCore APIs, identity controls, and observability tools. Agents can invoke one another as tools within a shared session, iterating autonomously until completion. The service supports multiple AI frameworks, including CrewAI, LangGraph, LlamaIndex, and Strands, and accommodates any model choice.

How it works

Runtime instances complement the existing runtime microVMs in AgentCore, which are optimized for shorter, stateless invocations. A typical deployment might use a lightweight orchestrator agent on microVMs to handle API calls and task routing, while worker agents on instances perform compute-heavy tasks like code compilation or security scanning. Both compute options are accessible through the same AgentCore runtime APIs.

To deploy agents on runtime instances, developers first create a capacity provider, which defines the underlying EC2 infrastructure. This involves selecting an operating system (Linux 64-bit ARM or x86_64), instance type (e.g., c7g.2xlarge for 8 vCPUs and 16 GiB memory), and configuring networking, storage, and security settings. Once the capacity provider is active, developers create a runtime, upload their agent code (as a zip file or container image), and specify the entry point. Agents sharing the same capacity provider run on the same EC2 infrastructure, enabling collaboration through a shared file system.

The AWS blog post demonstrates this workflow with two agents: a code writer and a code reviewer. Both agents operate within the same session, sharing a file system to exchange code without direct API calls. The writer generates Python code from natural language prompts, while the reviewer analyzes the output for bugs, security issues, and style improvements. This pattern can be extended to include additional agents, such as test runners or documentation generators, all working within the same shared directory.

Practical implications

Runtime instances address several challenges in deploying AI agents at scale. For teams running multi-step workflows that span hours or days, the service eliminates the need to build custom infrastructure for session persistence and state management. The ability to hibernate and resume sessions further reduces costs, while GPU support enables high-performance tasks like code compilation or security scanning. Integration with AgentCore Memory ensures agents retain context across sessions, improving continuity in long-running processes.

Pricing follows standard EC2 rates with an additional management fee for AgentCore orchestration. The service is available in multiple AWS regions, including US East, US West, Asia Pacific, and Europe. Developers can get started through the Amazon Bedrock AgentCore documentation or the AWS Management Console.