AWS has released HyperPod InstantStart, an open-source control plane that turns multi-step SageMaker HyperPod operations into guarded, retryable workflows. Infrastructure teams can use either a web interface or an AI agent; both call the same backend and read the same persisted operation state.
A request to create a cluster can span networking, accelerator capacity, dependencies, storage and identity. The agent plans those stages, launches each operation and waits for asynchronous AWS tasks to finish. It pauses for choices that remain the operator’s responsibility, such as Availability Zone, instance type and capacity type, rather than silently deciding them.
InstantStart runs as an out-of-band management container in the customer’s AWS account. It calls AWS and Kubernetes APIs but does not sit in the path of training jobs or inference requests. Everything it creates remains a standard resource that administrators can inspect with the AWS command-line tools or kubectl.
The boundary is important: Amazon EKS remains the user-managed orchestration layer, while HyperPod provides managed health checks, node recovery, capacity features, training recovery and inference capabilities. InstantStart composes those parts but does not eliminate the customer’s security and capacity responsibilities. By exposing a constrained Model Context Protocol toolset instead of a raw shell, the project offers a practical pattern for letting agents operate infrastructure without granting them an unlimited command surface.