이미지: METAL LAB 생성
Summary
- AWS has added a new compute option called "runtime instances" to Amazon Bedrock AgentCore
- Unlike existing microVMs (up to 8 hours), it maintains sessions for up to 14 days and supports GPU acceleration and multi-agent collaboration
- It supports frameworks like CrewAI, LangGraph, LlamaIndex, and Strands along with any model, and integrates with EBS and AgentCore Memory
- 발표
- Amazon Bedrock AgentCore Runtime instances
- 최대 세션 지속시간
- 14일
- 기존 마이크로VM 지속시간
- 최대 8시간
- GPU 지원
- 지원
- 지원 프레임워크
- CrewAI, LangGraph, LlamaIndex, Strands
- 연동 스토리지
- Amazon EBS, AgentCore Memory
- 데모 모델
- Claude Sonnet 4.5 (Bedrock 경유)
Always-on infrastructure is now needed for agents too
AWS has announced a new compute option called "runtime instances" for Amazon Bedrock AgentCore Runtime. The move is reportedly aimed at addressing the needs of AI agents moving beyond the prototype stage into production, which must maintain state across multi-day, multi-step workflows, share context with other agents, and access GPUs when necessary.
The existing AgentCore Runtime microVMs run for up to 8 hours and support stateful workflows through managed session storage. However, the company explained that tasks requiring continuous execution over multiple days, GPU or operating system access, or collaboration among multiple agents on the same host required a separate, larger-scale environment.
What runtime instances provide
Runtime instances allow multiple agents to be deployed to a single runtime on AWS-managed EC2-based infrastructure. Each agent can have its own independent dependencies and artifact types while collaborating on the same host within a shared session that persists for up to 14 days. The offering also supports GPU acceleration for compute-intensive tasks, session pause and resume to reduce costs during idle time, and containerized deployment for teams that want to deploy independently. Knowledge that needs to persist beyond a session is designed to be combined with Amazon Elastic Block Store (EBS) and AgentCore Memory to implement long-term memory.
What sets microVMs apart from instances
Both compute options share the same AgentCore Runtime API, identity controls, and observability tools, but they serve different roles.
| Category | Runtime microVM | Runtime instance |
|---|---|---|
| Maximum runtime | 8 hours | 14 days |
| GPU support | Not supported | Supported |
| Multi-agent collaboration | Limited | Supports collaboration on the same host |
| Best suited for | Rapid scaling, API calls, task routing | Tasks requiring persistent state, such as code compilation, security scanning, and GUI automation |
The two options can also be combined. In this structure, a lightweight orchestrator agent running on a microVM quickly handles API calls and task routing, while a worker agent running on an instance handles compute-intensive tasks that require persistent state and direct OS access.
Demonstrated with code-writing and code-review agents
AWS presented a demo featuring a code-writing agent that generates Python code from natural language descriptions, and a code-review agent that analyzes the generated code for bugs, security issues, and style. The two agents were configured to share the same file system, allowing the reviewer to read the writer's output directly without data transfers or API calls. The model used in the demo was reportedly Claude Sonnet 4.5, provided via Bedrock.
AWS developer Sébastien Stormacq explained that "packaging requires nothing more than the @app.entrypoint decorator and a single zip file."
Freedom to choose frameworks and models
A key feature of runtime instances is that they are designed to work as-is with a variety of agent frameworks—including CrewAI, LangGraph, LlamaIndex, and Strands—and with any model. This framework- and model-agnostic structure aligns with a broader trend among cloud providers emphasizing developer convenience in the competition over agent infrastructure. A similar trend can be seen in articles on cloud agent infrastructure trends covered by metallab.ai.
With this announcement, AWS appears to be positioning itself to capture the infrastructure layer needed as agents expand from short-term tasks to long-running autonomous workflows. However, concrete data on stability or cost structure in actual production environments has not yet been disclosed.
