AWS Bedrock AgentCore Adds Persistent EC2 Runtime for Long-Running Agents

Amazon Web Services has expanded its Bedrock AgentCore platform with runtime instances, a persistent compute option designed for production agents that exceed the eight-hour limit of its original serverless microVM architecture. The new capability keeps the same AgentCore APIs, security model and monitoring tooling while moving compute onto managed EC2 hardware in customer accounts.

The Problem: Longer-Running Agents Need Persistent Infrastructure#

AgentCore initially launched with serverless microVM sessions that handled short, request-response workloads efficiently. However, teams building complex multi-agent systems increasingly needed continuous execution, GPU acceleration, and tight collaboration between specialized agents—capabilities that the original eight-hour ceiling couldn't support. Runtime instances bridge this gap by offering sessions lasting up to fourteen days on EC2 infrastructure.

How Runtime Instances Work#

The feature operates through a capacity provider primitive that acts as a contract between agents and the EC2 resources AgentCore will provision and manage automatically. Teams define bounds like minimum and maximum instance counts, target utilization, and session lifespans, without having to wrestle with Auto Scaling groups or AMI pipelines directly.

Key capabilities include:

• Multi-agent co-location: Multiple agents run on the same host and share a session directory, letting them iterate autonomously without API calls for every handoff
• GPU and memory support: Instance types with GPU acceleration and large memory footprints now become viable
• Framework compatibility: Teams can deploy CrewAI, LangGraph, LlamaIndex, and Strands using simple packaging—just add an @app.entrypoint decorator and a zip file or container image
• File system persistence: Shared storage persists across agent interactions, critical for workflows managing code, tests, documentation and security artifacts

Complementary, Not Replacement#

AWS positions runtime instances as a companion to microVMs rather than a successor, enabling hybrid topologies where lightweight orchestrator agents run on microVMs and dispatch heavy lifting to instance-based workers. Short, bursty traffic remains cheaper on microVMs (billed per second with per-vCPU and per-GB pricing), while persistent or stateful workloads benefit from EC2 economics.