THE ESSENTIALS IN 30 SECONDS
- Agent skills in the google/skills and google-deepmind/science-skills repositories let an AI agent plan an HPC cluster and wait for human approval before deploying it.
- Domain skills cover OpenFOAM, Siemens STAR-CCM+, GROMACS, AlphaFold 3, Devito and Open Porous Media runs.
- AlphaEvolve on Cloud Batch uses a Cluster Toolkit blueprint to fan out hundreds of candidate evaluations in parallel; Google gives no preview or GA status.
Google Cloud has used the run-up to SC26, the supercomputing conference held in Chicago from 15 to 20 November, to push AI agents deeper into high-performance computing. In a blog post dated 14 November, HPC product managers Megan Gawlik and Annie Ma-Weaver describe agent skills for building and operating clusters, a way to run DeepMind's AlphaEvolve at scale on Cloud Batch, and an update on hardware.
Agents that plan clusters, then ask first
The skills sit on top of Cluster Toolkit, Google's tool for deploying HPC environments from declarative YAML blueprints. They are published in two GitHub repositories, google/skills and google-deepmind/science-skills, and Google says they rely on deterministic, self-describing tools checked by automated end-to-end tests.
When asked to create a cluster, an agent following the infrastructure skill audits compute and storage quotas, runs quick checks on regional capacity, picks a validated blueprint and then stops until a human approves. Only after that does it issue the deployment command. Once a cluster is running, domain skills guide the agent through fluid-dynamics jobs with OpenFOAM and Siemens STAR-CCM+, life-science pipelines with GROMACS and AlphaFold 3, and subsurface simulations with Devito and Open Porous Media. Google says agents work within IAM, Kueue quota and budget limits, with actions logged in Cloud Logging, and that a troubleshooting skill can clear stuck queues or node drains.
AlphaEvolve at batch scale
AlphaEvolve, the Gemini-based agent that evolves and scores code against a target metric, became generally available on Google Cloud in July. The new AlphaEvolve on Cloud Batch setup uses a single Cluster Toolkit blueprint to provision the controller, after which Batch runs hundreds of candidate evaluations concurrently, on single machines or multi-node MPI groups. Researchers can steer runs from hosted Python notebooks, and evaluation can use Cloud RDMA, Managed Lustre, Spot VMs or Dynamic Workload Scheduler. The post does not say whether this setup is in preview or generally available.
Hardware and open questions
On infrastructure, Google lists A5X bare-metal instances on Nvidia's Vera Rubin NVL72 platform as newly available, alongside TPU 8t for large-scale training and simulation and TPU 8i for low-latency inference. Batch can now run jobs on single-host TPU slices, and Managed Lustre is quoted at up to 10 TB/s of read throughput. No performance figures are given for A5X or the TPUs, and no customers are named for the new skills, so how well agents handle real production clusters remains to be seen.
