AWS adds persistent compute option for production AI agents in Amazon Bedrock AgentCore

AWS announced runtime instances, a new compute option within Amazon Bedrock AgentCore designed to support production-grade AI agents. Unlike runtime microVMs— which handle stateful workflows for up to 8 hours—runtime instances provide managed EC2 infrastructure for longer-running, stateful agent workloads.

The new offering enables multiple agents to run on the same host within shared sessions that persist for up to 14 days. It supports GPU acceleration for compute-intensive tasks and allows agents to collaborate within the same runtime environment.

Runtime instances also include features such as session stop/restart to reduce costs during idle periods and containerized deployments. These capabilities are intended to simplify the operational complexity of deploying AI agents in production, particularly for workflows that require persistent state, multi-day execution, or specialized hardware like GPUs.

The announcement highlights the challenges of moving AI agents from prototype to production, where infrastructure demands increase due to the need for state persistence, multi-step coordination, and access to specialized compute resources.