HPC expertise is packaged as executable guidance
Google Cloud has introduced reusable agent skills for high-performance computing and a cloud-scale AlphaEvolve solution at SC26. The release packages infrastructure and scientific expertise into guided workflows that can validate quotas, select tested Cluster Toolkit blueprints, deploy clusters and diagnose operational problems. It also lets research teams use a Gemini-powered evolutionary loop to mutate and benchmark algorithms across large batches of cloud resources. The goal is not an unguided agent with shell access, but repeatable automation surrounded by deterministic tools and approvals.
HPC environments are difficult for general-purpose agents because small configuration errors have expensive consequences. A model can select the wrong storage tier, ignore placement constraints or provision a cluster that cannot run a tightly coupled workload efficiently. Google’s new skills describe the required sequence and connect the agent to self-describing tools. The templates are published through Google Skills and DeepMind Science Skills repositories and are validated with automated end-to-end testing rather than presented only as prompt examples.
Infrastructure skills begin by auditing compute and storage quotas and checking regional capacity. They then match the workload to an established Cluster Toolkit blueprint and wait for human approval before running deployment. This sequencing is important. It separates the model’s role in interpreting intent from the controls that decide whether resources may be created. Organisations should preserve that separation when adapting a skill, particularly where a single request could reserve costly accelerators or alter a shared research environment.
Human approval stays before deployment
Workload-specific skills add scientific context after the cluster exists. Google lists workflows for OpenFOAM and Siemens STAR-CCM+ fluid dynamics, GROMACS and AlphaFold 3 life sciences and Devito and Open Porous Media subsurface simulations. These skills can encode command structure, expected artefacts and troubleshooting knowledge that a general assistant would otherwise reconstruct imperfectly. Scientists still need to verify assumptions, boundary conditions and results; an automated run does not confer scientific validity.
Governance extends into day-two operations through IAM, Kueue quotas, budgets, Cloud Logging and troubleshooting skills for problems such as stuck queues and drained nodes. This is where agentic operations can earn trust. Provisioning a cluster once is a demonstration; recovering it predictably under resource pressure is an operational capability. Teams should test cancellation, partial deployment, quota exhaustion and unavailable capacity, then confirm that the agent produces an understandable record rather than hiding recovery actions inside a long conversation.
AlphaEvolve addresses a different part of the scientific loop. Developed by Google DeepMind, it uses Gemini to propose mutations to code or engineering parameters, evaluates candidates and repeats against defined metrics. The Cloud Batch solution fans out hundreds of evaluations across isolated virtual machines or multi-node MPI groups. Researchers can steer experiments from notebooks or command-line tools and inspect candidate lineage instead of accepting a final answer without evidence of how it emerged.
AlphaEvolve scales algorithm search
Evaluating candidates on production-like networks and storage can prevent an optimisation that performs well only on a laptop. Google supports Cloud RDMA and Managed Lustre, with Spot VMs or Dynamic Workload Scheduler used to balance cost and completion time. The evaluation objective still deserves scrutiny. If the metric rewards speed alone, the agent may trade away accuracy, reproducibility or resource efficiency. Scientific teams should use multiple acceptance tests and retain the code, environment and data needed to reproduce a winning candidate.
The infrastructure announcement spans H4D and C4A compute, Blackwell and Vera Rubin GPU systems, eighth-generation TPUs and Managed Lustre. Google says the storage platform can provide up to 10 TB per second of read throughput and an 80 PB dynamic tier. These figures describe available building blocks, not a guaranteed application result. Network topology, data layout, software licences and queue availability will determine whether an agent-designed environment actually improves time to insight.
For Australian research and engineering organisations, region availability and data movement will be as important as global product specifications. A skill may choose a technically ideal accelerator that is unavailable locally or requires data to cross a governance boundary. Platform teams should constrain eligible regions, machine families and budgets before the agent plans. They also need a graceful fallback that explains why a preferred design cannot be deployed rather than silently substituting a materially different architecture.
Infrastructure still determines useful autonomy
Google’s SC26 release points towards a more disciplined version of agentic infrastructure: expert procedures are versioned, tools are deterministic and high-impact changes wait for approval. AlphaEvolve adds an exploratory layer where models can search a large design space, but evidence comes from real benchmarks. The combination could make specialist HPC knowledge more accessible. Teams should record the precise skill, blueprint, model and benchmark environment behind every accepted optimisation so later researchers can reproduce or challenge the result. Its success will depend on whether teams keep scientific judgement, cost authority and reproducibility outside the model’s unilateral control.