Runtime Instances: Persistent Computing for Production AI Agents on Amazon Bedrock AgentCore

Amazon Bedrock AgentCore Runtime Instances is an AWS-managed persistent computing option for complex production AI agent workloads. It allows deploying multiple agents on AWS-managed EC2 infrastructure and supports their collaboration on the same host and in shared sessions.

Runtime Instances provide shared sessions lasting up to 14 days, GPU acceleration, container deployment, and session stop/resume capabilities. Long-idle workloads can hibernate their environment and later resume with state preserved.

The existing AgentCore Runtime microVM is suitable for invocations running up to 8 hours and fast scaling, while Runtime Instances are suitable for tasks lasting several days or requiring GPU, direct OS access, and multi-agent collaboration. Both environments use the same AgentCore Runtime API, identity controls, and observability features, so they can be configured together.

In a representative configuration, an orchestrator agent running on a microVM distributes tasks and aggregates results, while worker agents on Runtime Instances handle tasks requiring persistent state and direct OS access, such as code compilation, security scanning, and GUI automation.

Agents can use any desired framework and model, including CrewAI, LangGraph, LlamaIndex, and Strands. Deployment requires only the @app.entrypoint decorator and a ZIP file or container image, and data requiring long-term retention can utilize Amazon EBS and AgentCore Memory together.

In the introduced example, a code-writing agent and a code-review agent share the same file system. The code-writing agent generates files, and the review agent reads and analyzes them without separate data transfers or API calls.