Google Cloud adds agent skills to Cluster Toolkit for HPC workflows
Google Cloud introduced reusable agent skills that guide Cluster Toolkit deployments and scientific workloads through capacity checks, approvals, governance and diagnostics.
Google Cloud introduced reusable, composable agent skills at SC26 to guide AI agents through high-performance computing deployments and scientific workloads built with its Cluster Toolkit.
According to Google Cloud, the infrastructure workflow checks compute and storage quotas, runs regional-capacity canaries and selects a validated blueprint. Human approval is required before the agent can deploy a cluster with gcluster.
Cluster Toolkit is Google Cloud’s open-source framework for deploying HPC environments. Users define an environment in a declarative YAML blueprint assembled from reusable modules, then use the gcluster command-line tool to generate a deployment folder with Terraform or Packer configuration.
Google Cloud said the skills use self-describing, deterministic tools and undergo automated end-to-end testing. They are available through the public Google Skills repository and Google DeepMind Science Skills repository. For context, Google DeepMind previously released Antigravity cheminformatics skills that also combine agent orchestration with deterministic tools.
The workload-specific examples span several research domains. Google Cloud said the skills can guide OpenFOAM and Siemens STAR-CCM+ fluid-dynamics jobs, GROMACS and AlphaFold 3 life-sciences pipelines, and Devito and Open Porous Media subsurface simulations.
Google Cloud also said agents can operate within Identity and Access Management controls, Kueue quotas and budget guardrails while recording provenance in Cloud Logging. A gke-kueue-debugging skill is intended to diagnose stuck queues and node drains.
The Cluster Toolkit source repository includes context files that direct compatible AI coding assistants to a cluster-blueprint-assistant skill. Google Cloud’s documentation says those files are not included in the prebuilt release bundle.
Before any infrastructure is provisioned, gcluster explain can list available blueprints and modules, show required APIs and variables, and match a workload specification to a blueprint. Google Cloud says this inspection runs offline.
The announcement did not identify a versioned release or commit mapping every named HPC workflow to a downloadable skill. The opened sources also provided no independent benchmarks for setup time, error rates, scientific productivity or total cost, and did not establish production customer adoption.
More news

NVIDIA Introduces Four-Node DGX Spark HPC Developer Kit

NVIDIA Specifies 192-CPU Vera Rack for Scientific Computing

AMD-powered Lux supercomputer opens to Oak Ridge researchers
