Google Wants AI Agents to Run the Supercomputer: Cloud Brings HPC 'Skills' and AlphaEvolve on Batch to SC26
As the SC26 supercomputing conference opens in Chicago, Google Cloud is pitching agents that check quotas, pick validated cluster blueprints and wait for human approval before deploying, plus a packaged way to run DeepMind's AlphaEvolve code optimizer across hundreds of cloud machines at once.
Published
Last updated / Fact-checked on

Illustrative photo of networking cables in a server rack. It does not show Google Cloud hardware.
Supercomputing has always been a field for specialists. Setting up a cluster, choosing the right storage and network, and getting a physics code to scale across hundreds of machines can take weeks of expert work. Google Cloud's bet for SC26, the supercomputing conference that runs Nov. 15 to 20 in Chicago, is that AI agents can do much of that work, if they are taught the rules first.
In a blog post dated Nov. 13, product managers Megan Gawlik and Annie Ma-Weaver laid out three pieces: reusable "skills" that teach agents how to build and run high-performance computing (HPC) workloads, a packaged way to run Google DeepMind's AlphaEvolve optimizer at scale on Cloud Batch, and the hardware lineup underneath.
The pitch comes with a warning about general-purpose agents. Without deep context, the authors write, 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."
At a glance
| Item | Detail | Source |
|---|---|---|
| Event | SC26, Nov. 15 to 20, McCormick Place West, Chicago | Google Cloud Blog |
| Agent skills | Infrastructure, workload and diagnostic skills built on Cluster Toolkit | Google Cloud Blog |
| Where to get them | Google Skills and DeepMind Science Skills repositories | Google Cloud Blog |
| Optimizer | AlphaEvolve on Cloud Batch, hundreds of parallel evaluations | Google Cloud Blog, docs |
| New hardware note | A5X bare metal on Nvidia Vera Rubin NVL72 "newly available" | Google Cloud Blog |
| Storage claim | Managed Lustre up to 10 TB/s read throughput | Google Cloud Blog |
| Pricing | Not given | — |
Skills that check before they deploy
Google already offers Cluster Toolkit, which builds repeatable HPC environments from YAML "blueprints." The new layer pairs those blueprints with agent skills, which Google describes as self-describing, deterministic tools tested end to end.
The infrastructure skills are built to be cautious. When a user asks an agent to stand up a cluster, Google says the skill makes it first audit compute and storage quotas, run quick checks for regional capacity, match the workload to a validated blueprint and "wait for human approval" before running the deploy command.
Workload skills cover specific science and engineering software. 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.
Google also says the agents work inside IAM permissions, Kueue quotas and budget limits, log their actions to Cloud Logging, and use troubleshooting skills such as one named gke-kueue-debugging to clear stuck queues or node drains.
AlphaEvolve, scaled out
The second piece is AlphaEvolve, the Google DeepMind system that uses a Gemini-powered evolutionary loop to mutate, benchmark and improve code against a target metric. The blog says it can also tune 3D engineering parameters.
The new "AlphaEvolve on Cloud Batch" solution is meant to take an algorithm from a laptop to large-scale testing. A single Cluster Toolkit blueprint sets up the controller, and Batch then fans out "hundreds of candidate evaluations concurrently," either in single-machine sandboxes or multi-node MPI jobs. Researchers can steer runs from hosted notebooks or plug them into command-line and agent tools.
Google's documentation, updated Nov. 13, describes the loop in four steps: a controller asks the Gemini-backed AlphaEvolve API for candidate code, Batch jobs test each candidate for correctness and speed, results are stored in a Cloud Storage bucket, and the results are fed back to guide the next round. Because the artifacts live in storage, users can tear down the compute when it is idle.
The authors argue that where code is tested matters. Evolving an algorithm on one node with slow networking "risks producing code that only performs well at small scale," they write. Testing candidates on machines with Cloud RDMA networking and Managed Lustre storage is meant to find code tuned for real supercomputer fabrics. Users can use Spot VMs or Google's Dynamic Workload Scheduler to trade cost against speed.
Google names the industries it expects to care, including financial services, chip design, manufacturing, logistics, pharmaceuticals and AI training. It does not name customers or share results from them.
The hardware underneath
The post groups the infrastructure under Cluster Director. It lists H4D instances with 5th Gen AMD EPYC processors, including 3 TB high-memory shapes, Arm-based C4A instances on Google Axion, A4 and A4X instances with Nvidia Blackwell B200 and GB200 NVL72, and G4 instances with Nvidia RTX PRO 6000 Blackwell GPUs.
It calls A5X bare-metal instances, built on Nvidia's Vera Rubin NVL72 platform, "newly available." It also highlights Google's eighth-generation TPUs, TPU 8t for training and simulation and TPU 8i for low-latency inference, and says Batch now supports jobs on single-host TPU slices. For storage, Google claims Managed Lustre can deliver up to 10 TB/s of read throughput, with GPUDirect Storage over RDMA and an 80 PB Dynamic Tier.
Why it matters
HPC users are among the earliest adopters of new computing ideas, as the post itself notes, and they are a demanding test for AI agents. A wrong storage choice or a badly placed job can waste large sums on idle machines. Google's answer is to wrap agents in guardrails: validated templates, quota checks, budget limits and a human sign-off before deployment. That is a different approach from letting a general model loose in a terminal, and the blog says so: agentic HPC "is not about turning unguided models loose on a terminal."
AlphaEvolve on Batch also points to where cloud providers see value. Faster code lowers the cost of every future run, so automated optimization at scale could matter more over time than any single hardware launch.
What to watch
The real test is whether research labs and companies report concrete gains, such as faster codes or shorter setup times, from these tools. Watch for pricing, for named customers, and for how the agent skills handle failures on real clusters. Google says it will demo Antigravity for HPC and AlphaEvolve for HPC at booth #3418 at SC26, and rival clouds are likely to show their own agent tools at the same event.
How we reported this
DailyMech read Google Cloud's SC26 blog post and the AlphaEvolve for HPC documentation page in full. The blog post is dated Nov. 13; we are covering it now, as SC26 opens, because DailyMech had not reported on it. The post does not describe AlphaEvolve on Cloud Batch as "generally available," so we do not either. Performance figures are Google's own. DailyMech did not test the tools and contacted no one for this article.
Published November 15, 2026. Facts checked against the cited sources by DailyMech's AI writer.
Sources
- Where supercomputing meets agentic AI: What's new in Google Cloud HPC and AI at SC26, Megan Gawlik and Annie Ma-Weaver, Google Cloud Blog, Nov. 13, 2026 (fully read).
- AlphaEvolve for HPC use cases, Google Cloud documentation, last updated Nov. 13, 2026 (fully read).
DailyMech used AI tools to research and draft this article. Facts were checked against the cited sources by DailyMech's AI writer, not by a human editor.