Standing up a training cluster on AWS may soon be a sentence typed to an agent rather than a checklist clicked through a console. The company open-sourced InstantStart, a control plane that joins Kubernetes orchestration with SageMaker HyperPod’s managed features, and detailed the project in a machine learning blog post this week.
InstantStart lives as one container inside a customer account, out of the data path of training and inference traffic, speaking to AWS service APIs and the Kubernetes API. Its web interface, REST API, and Model Context Protocol tools all sit behind the same backend, so a validation written once protects every entry point. In the browser, cluster creation is a form with a progress panel; from a terminal, the same task is a natural-language request to hypd-inst-agent, an agent configuration built for Kiro CLI.
The agent sequences the operation in stages: spinning up the EKS control plane, which AWS puts at roughly 8 to 12 minutes, picking a cluster, reconciling dependencies, creating the HyperPod cluster with automatic node recovery enabled, and mounting storage. Each stage records its own status and can retry independently. Skills in the repo encode three rules: poll long-running operations until they finish, ask only decision-grade questions like availability zone and instance type, and inspect what already exists before creating anything. Every resource produced is a standard AWS or Kubernetes object, visible through the CLI and kubectl.
The release is another sign that infrastructure operators expect agents to run fleets directly, with MCP emerging as the common language between a model and the control plane.
