Amazon Bedrock AgentCore launches runtime instances for persistent AI agent compute

Amazon Web Services announced that its Bedrock AgentCore platform now offers "runtime instances," a new compute option that lets customers run AI agents on managed Amazon EC2 instances. The service provides dedicated, long‑running infrastructure—supporting sessions up to 14 days, GPU acceleration, and a range of instance types (memory‑optimized, compute‑optimized, and GPU families). Runtime instances handle provisioning, patching, scaling, and lifecycle management, and can be mixed with the existing microVM‑based runtime that supports shorter, up‑to‑8‑hour sessions.

The rollout also adds new capabilities for users in AWS GovCloud (US‑West), including short‑term and long‑term memory for agents, fine‑grained policy controls expressed in natural language and compiled to the Cedar policy language, and a managed harness that simplifies model, tool, and instruction configuration. The features are available in multiple AWS regions worldwide, and customers are billed for the underlying EC2 usage in addition to standard AgentCore fees.