Google Cloud Gives AI Agents HPC Skills
Google Cloud said on Nov. 14, ahead of SC26 in Chicago, that it has paired Cluster Toolkit with reusable agent skills from its Google Skills and DeepMind Science Skills repositories. The skills audit quotas, check capacity and wait for human approval before deploying, and guide agents through codes such as OpenFOAM, GROMACS and AlphaFold 3. Google also launched AlphaEvolve on Cloud Batch.
Google Cloud is giving AI agents packaged know-how for running supercomputing jobs. In a blog post dated Nov. 14 ahead of the SC26 supercomputing conference in Chicago, HPC product managers Megan Gawlik and Annie Ma-Weaver said the company has paired its Cluster Toolkit with reusable agent skills that set up clusters, configure scientific software and pause for human approval before deploying.

Key takeaways
- Agent skills for Google Cloud's Cluster Toolkit are downloadable from two Apache 2.0 GitHub repositories, Google Skills and DeepMind Science Skills.
- Infrastructure skills audit quotas, run capacity checks and wait for human approval before running the deploy command.
- Google also launched AlphaEvolve on Cloud Batch, which fans out hundreds of candidate code evaluations in parallel.
What are Google Cloud's agentic HPC skills?
Cluster Toolkit already lets teams deploy HPC environments from declarative YAML blueprints. The new layer adds what Google calls composable agent skills: self-describing, deterministic tools validated through automated end-to-end testing.
The problem Google describes is that a general-purpose model without infrastructure context makes costly mistakes. The post says an unguided agent can pair incompatible storage tiers with compute-optimized VMs, overlook placement policies or stall when diagnosing a multi-node MPI fabric error.
When asked to stand up a cluster, the infrastructure skill directs the agent to audit compute and storage quotas, run canary checks for regional capacity, match the workload to a validated blueprint and wait for human approval before running gcluster deploy. Agents operate inside IAM, Kueue quota and budget guardrails, with provenance logged in Cloud Logging.
Which scientific workloads are covered?
Once a cluster is live, domain skills guide the agent through specific codes. Google lists OpenFOAM and Siemens STAR-CCM+ for fluid dynamics, GROMACS and AlphaFold 3 for life sciences, and Devito and Open Porous Media for subsurface simulation with browser-based 3D visualization.
The DeepMind Science Skills repository describes itself as a collection spanning genomics, structural biology, cheminformatics and literature search, integrating AlphaGenome, the AlphaFold Protein Structure Database, UniProt and more than 30 other databases and tools. It installs with a single npx skills add command. Some skills, such as AlphaGenome, require an API key.
AlphaEvolve on Cloud Batch
Google also announced an AlphaEvolve on Cloud Batch solution. AlphaEvolve, developed by Google DeepMind, uses a Gemini-powered evolutionary loop to mutate, benchmark and optimize code against target metrics. A single blueprint provisions the controller, and Batch runs hundreds of candidate evaluations concurrently across single-VM sandboxes or multi-node MPI task groups, using Spot VMs or Dynamic Workload Scheduler to balance cost and speed.
Context: hardware under the agents
The post lists the infrastructure the skills target: H4D instances on 5th Gen AMD EPYC processors with 3 TB high-memory shapes, newly available A5X bare-metal instances on NVIDIA's Vera Rubin NVL72 platform, eighth-generation TPU 8t and 8i chips, and Managed Lustre storage delivering up to 10 TB/s of read throughput. Google's push to package expertise for agents follows DeepMind tools such as WeatherNext's cyclone forecasts.
"Agentic HPC is not about turning unguided models loose on a terminal," Gawlik and Ma-Weaver wrote.
What's next
SC26 runs Nov. 15–20 at McCormick Place. Google says it will demo Antigravity for HPC, an HPC Prompt Playground and AlphaEvolve for HPC at Booth #3418. Google did not publish benchmarks showing how often skill-guided agents avoid deployment errors, so independent results from research computing centers will be the test.
FAQ
Do the skills deploy clusters automatically? The infrastructure skill prepares and validates a deployment, but Google says it waits for human approval before running the deploy command.
Are the skills free? Both repositories are public on GitHub, and their software is licensed under Apache 2.0. Running clusters on Google Cloud is billed as usual.
Which agents can use them? The Google Skills repository lists plugin installs for Claude Code, Codex and the Antigravity CLI; Science Skills can be installed in Google Antigravity.
References
- Where supercomputing meets agentic AI: What's new in Google Cloud HPC and AI at SC26 — Google Cloud (Nov. 14, 2026)
- google-deepmind/science-skills — Google DeepMind, GitHub
- google/skills: Agent Skills for Google products and technologies — Google, GitHub
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.