Amazon Bedrock AgentCore adds runtime instances for production AI agents

Runtime instances provide AWS-managed EC2 infrastructure for Amazon Bedrock AgentCore, allowing multiple agents to run in a single environment with persistent sessions up to 14 days. This feature supports GPU acceleration and containerized deployments, simplifying the management of complex agent workloads that previously required self-provisioned infrastructure.

Developers can package agents as zip files or container images, bringing their own frameworks like CrewAI, LangGraph, or LlamaIndex. The runtime instances simplify deployment with minimal packaging requirements, such as an @app.entrypoint decorator.

Agents deployed on runtime instances can now call each other as tools within shared sessions, facilitating autonomous iteration and complex task completion. Session storage persists for up to 14 days, and agents can be hibernated and resumed, preserving state across extended periods.

Runtime instances integrate with Amazon Elastic Block Store (EBS) and AgentCore Memory, providing agents with long-term recall capabilities that extend beyond individual sessions and environments.

Runtime instances and existing runtime microVMs are complementary. A lightweight orchestrator on microVMs can dispatch tasks to specialized worker agents running on instances, leveraging the strengths of both compute options for efficient workflow management.

A demo showcases two agents, a code writer and a code reviewer, collaborating within a shared file system on the same runtime instance. The example details the steps for creating a capacity provider, deploying agents via zip files, and invoking them through the AWS Management Console.