Computational scientists and HPC engineers have always been among the earliest adopters of new computing paradigms, and agents are no exception. Today, research and engineering teams use AI models and agents for things such as ingesting dense scientific literature, optimizing code, and autonomously steering simulations.
Yet moving from an agent that suggests code optimizations to an agentic system that reliably architects, provisions, and optimizes production HPC workloads at scale takes more than a general-purpose reasoning model. Without deep infrastructure and domain context, an unguided agent can still pair incompatible storage tiers with compute-optimized VMs, overlook placement policies, or stall when diagnosing a multi-node MPI fabric error.
That challenge led our team to ask a simple question: What if we could package our deep HPC and scientific expertise into reusable agent skills, turning complex setup and optimization into guided workflows?
At SC26, we are showing how Google Cloud brings agentic HPC to life with reusable skills that teach agents how to build and run HPC workloads. In addition, we’re also demonstrating evolutionary agents that optimize your algorithms at cloud scale, and purpose-built supercomputing infrastructure designed for where HPC meets AI.
Organizations rely on Cluster Toolkit to deploy repeatable, production-ready HPC environments on Google Cloud through declarative YAML blueprints. Now we are expanding these capabilities by pairing Cluster Toolkit with composable agent skills with self-describing, deterministic tools, validated through automated, end-to-end testing and available for download from the Google Skills and DeepMind Science Skills repositories. Together, this delivers:
Infrastructure skills that validate before they deploy: When you ask an agent to stand up a cluster, the skill directs it to first audit your compute and storage quotas, run rapid canary checks for regional capacity, match your workload to a validated Cluster Toolkit blueprint, and wait for human approval, all before running the gcluster deploy command.
Workload-specific skills that know the science: After your cluster is live, domain skills connected to Google Skills and DeepMind Science Skills guide the agent through things like configuring OpenFOAM and Siemens STAR-CCM+ fluid dynamics runs, tuning GROMACS and AlphaFold 3 life-sciences pipelines, or launching Devito and Open Porous Media (OPM) subsurface simulations with browser-based 3D visualization.
Built-in governance and Day-2 diagnostics: Agents operate inside strict IAM, Kueue quota, and budget guardrails, with full provenance in Cloud Logging and structured troubleshooting skills (gke-kueue-debugging) to resolve stuck queues or node drains automatically.
Configuring and running existing solvers is only part of the discovery loop. Across financial services, electronic design automation (EDA), manufacturing, logistics, pharmaceuticals, and AI training, competitive advantage comes from having the fastest underlying code and algorithms. Historically, squeezing out performance gains required scarce optimization specialists and costly brute-force compute.
Developed by Google DeepMind, AlphaEvolve changes the equation by using a Gemini-powered evolutionary loop to iteratively mutate, benchmark, and optimize code and 3D engineering parameters against target metrics. The AlphaEvolve on Cloud Batch solution, streamlines the process of moving from a seed algorithm on a laptop to massive parallel experimentation. With Batch, you get:
Automated parallel evaluation: A single gcluster blueprint provisions the AlphaEvolve controller environment and uses Batch to fan out hundreds of candidate evaluations concurrently across single-VM sandboxes or multi-node MPI task groups.
Interactive control from your preferred surface: Researchers can steer evolutionary experiments and inspect candidate lineage through hosted Python notebooks, or integrate optimization runs directly into their existing CLI workloads and agent-driven development tools.
Evaluation at production scale: Evolving an algorithm on a single node with slow networking risks producing code that only performs well at small scale. By evaluating candidates across Google Cloud compute instances with Cloud RDMA and Google Cloud Managed Lustre — and using Spot VMs or Dynamic Workload Scheduler (DWS) to balance cost against turnaround time — you discover algorithms tuned for real supercomputing fabrics, permanently lowering the compute cost of every future production run.
An autonomous workflow is only as capable as the infrastructure beneath it. As agentic pipelines pivot dynamically between memory-bandwidth-bound physics solvers, domain foundation models (like AlphaFold 3, AlphaGenome, and WeatherNext), and large-scale AI training, Google Cloud provides a unified hardware and orchestration fabric managed by Cluster Director that offers:
Tightly coupled HPC and scale-out compute: Run classic MPI simulations on ** H4D** instances powered by 5th Gen AMD EPYC processors and Cloud RDMA, including 3 TB high-memory shapes built for large EDA and engineering models. Alternatively, you can maximize performance-per-watt on Arm-based C4A instances powered by Google Axion.
Next-generation GPU acceleration: Run coupled physics-AI pipelines and multi-node GPU simulations on A4 and A4X instances (NVIDIA Blackwell B200 and GB200 NVL72), scale frontier AI training and reasoning on newly available A5X bare metal instances built on the NVIDIA Vera Rubin NVL72 platform, and use G4 instances with NVIDIA RTX PRO 6000 Blackwell GPUs to accelerate CAE, CFD, and 3D visualization while reducing per-core software licensing costs.
8th gen TPUs for the agentic era: Deploy 8th generation TPUs, featuring TPU 8t for large-scale pre-training, reinforcement learning, and accelerated mathematical modeling and simulation (ModSim), alongside TPU 8i for high-throughput, low-latency reasoning and agentic inference. Additionally, Batch now supports running jobs on single-host TPU slices, providing flexible provisioning options for accelerated HPC and AI/ML workloads.
Sub-millisecond parallel storage: Reduce data bottlenecks between simulation and AI stages with Google Cloud Managed Lustre, delivering up to 10 TB/s of read throughput, GPUDirect Storage over RDMA, and an 80 PB Dynamic Tier with Cloud Storage Mirroring. For even more performance, you can pair it with Cloud Storage FUSE local SSD caching.
Agentic HPC is not about turning unguided models loose on a terminal. This shift is about combining the reasoning power of Gemini with the codified engineering of Cluster Toolkit skills, evolutionary optimization on Batch, and purpose-built supercomputing hardware — so your teams spend their time on discovery rather than infrastructure plumbing.
Visiting SC26 (November 15–20)? Stop by Booth #3418 at McCormick Place West for live demos of Antigravity for HPC, the HPC Prompt Playground, and AlphaEvolve for HPC, and join us at our Advanced Computing Community Half-Day Workshop and Agentic and Hybrid HPC Birds of a Feather (BoF) session.
To start building today, explore the Google Cloud Cluster Toolkit and Google Cloud HPC solutions. To access and implement these agentic workflows, download the validated templates from the Google Skills and DeepMind Science Skills repositories.