Next upHack for Humanity: San Francisco (powered by Google Gemini)
News

AWS introduces HyperPod InstantStart control plane for agent-driven cluster operations

AWS introduced HyperPod InstantStart, an open-source control plane that routes web and agent-driven SageMaker HyperPod operations through shared APIs, validations and persisted state.

D
Sep 4, 2026 · 1 min read

AWS introduced HyperPod InstantStart, an open-source control plane for agent-driven Amazon SageMaker HyperPod operations. Its web interface and its agent tools share one set of backend APIs, validations and persisted operation state instead of wrapping command-line or software development kit access directly.

InstantStart runs as an out-of-band management container inside a customer’s AWS account. It calls AWS service and Kubernetes APIs but does not sit in the path of training jobs or inference requests. Alongside that web interface and a set of REST APIs, the container exposes a Model Context Protocol server with 38 bounded tools covering cluster lifecycle, instance groups, managed features, storage, model downloads, inference deployments, jobs and node operations.

Setup is split into stages that include creating an Amazon EKS control plane, selecting an active cluster, reconciling dependencies, creating the HyperPod cluster and configuring storage. AWS said later stages persist their status and can be retried independently. Each operation records its phase before polling begins, and API responses determine terminal success or failure. That design, AWS said, prevents browser refreshes or agent retries from replaying mutations.

AWS also described field normalization and explicit-diff handling that preserve immutable or easily lost instance-group settings and keep stale interface values from being resubmitted during unrelated updates. Web and agent requests therefore run through the same validation and reconciliation paths, and the AWS and Kubernetes resources they create stay inspectable with the AWS CLI and kubectl.

The public repository documents workflows for cluster provisioning, capacity and resilience, training, inference, model downloads, storage and observability. AWS cautioned that teams still have to govern EKS security, agent credentials, workflow rules, costs and quotas, and to review unsupported feature combinations.

More news