AWS has announced runtime instances, a new compute option for Amazon Bedrock AgentCore Runtime, designed for production AI agents that need to stay active over long periods. The service provides AWS-managed EC2 infrastructure where multiple agents can be deployed in a single runtime, each with its own dependencies and artifact types. Agents can share sessions that persist for up to 14 days, and the service supports GPU acceleration, session stop/restart to save costs during idle time, and containerized deployments. It also works with Amazon EBS and AgentCore Memory for knowledge that needs to survive beyond a session.

Before this option, teams that wanted agents to run for days, access GPUs, or coordinate multiple agents on one host had to provision and manage EC2 instances, networking, session management, scaling, and monitoring themselves. Runtime instances handles that infrastructure while integrating with the same AgentCore APIs, identity controls, and observability used by AgentCore Runtime microVMs. Agents can call each other as tools within a shared session, and developers can bring any framework — such as CrewAI, LangGraph, LlamaIndex, or Strands — and any model. Packaging is minimal: an @app.entrypoint decorator plus a zip file or container image.

Runtime microVMs and runtime instances are complementary. A lightweight orchestrator on a microVM can coordinate and dispatch work to specialized workers running on instances, which handle compute-intensive tasks like code compilation, security scanning, or GUI automation that require persistent state and direct OS access. The existing microVMs support invocations up to 8 hours and stateful workflows through managed session storage.

In a demonstration, a code writer agent generated Python code and a code reviewer agent analyzed it. Both agents shared the same file system, so the reviewer read the file the writer created without any data transfer or API calls between them. Connecting them meant reusing the same session ID. The reviewer reported no critical bugs but suggested type hints, input validation, and simpler edge-case handling.

Key details include Linux support on ARM64 and x86_64, session persistence up to 14 days, Python 3.11–14 runtime support with native code, container images, GPU-accelerated instance types, and the same AgentCore API and control plane integrations. Pricing is standard EC2 pricing plus a management fee for AgentCore orchestration. Launched regions include US East (Ohio, N. Virginia), US West (Oregon), Asia Pacific (Mumbai, Singapore, Sydney, Tokyo), and Europe (Frankfurt, Ireland).

Source: AWS