Agentic HPC Console
active buildA console where research assistants describe the analysis they want and get a reviewable execution plan — sbatch arrays, partitions, cost estimate — before anything touches the cluster. The agent has a deliberately narrow tool surface, and every job it runs leaves provenance behind.
problem
Research assistants need HPC, not HPC expertise. The gap between "run this pipeline on batch 2026-07" and a correct sbatch array — right partition, right GPU count, scratch quota, module versions — is where their work stalls and where the cluster gets misused. Handing them a shell teaches Slurm; handing them a ticket queue teaches waiting.
approach
- Plan, then approve. The agent never submits directly. It emits a concrete plan — command, resources, runtime and cost estimate — that a human approves, edits, or cancels.
- Narrow tools, not shell access. Six typed tools over the Slurm REST API and AWS Batch: submit, status, tail, cancel, fetch results, describe partitions. No arbitrary execution path.
- Provenance by default. Every run records the prompt, the plan, the resolved script, container digests, and inputs — so a result can be traced back to what produced it.
- Accounting built in. SlurmDBD plus Aurora gives per-team chargeback and capacity planning, which is what makes self-service acceptable to the people who own the budget.
outcome
Assistants launch, monitor, and retrieve their own multi-hundred-task jobs without an engineer in the loop, and the approval step catches the expensive mistakes — wrong partition, missing array bounds, oversized GPU requests — before they queue.
> run the CNV pipeline on the 48 samples from batch 2026-07plan sbatch cnv.sbatch --array=1-48 --partition=gpu --gres=gpu:1 est. 3h20m · 48 tasks · ~$18.40 (spot) · fsx scratch 1.2 TB [approve] [edit] [cancel]✓ submitted job 481203 · 48 tasks queued◍ running 31/48 complete · 4 gpu nodes · 1h02m elapsed✓ provenance run manifest → s3://mbbg-runs/481203/manifest.jsonscreens
- Plan approval view
- Queue / job monitor