Skip to content

About

Participant kit from the NCSA workshop session "Getting Started Using AI Agents for Research/HPC" on Delta (Oct 2026): agent rules, session guides, prepared topics and helper scripts

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Delta agents workshop kit

The participant kit from the session "Getting Started Using AI Agents for Research/HPC" at the NCSA Regional Workshop on AI (Mon 2026-10-05), run on NCSA's Delta supercomputer. About 50 researchers each worked with a coding agent on a compute node: first a prepared topic taken to a checked result, then an open-ended problem of their own.

This repository is the kit as published to participants that morning. It is shared as a worked example of how to scaffold an agent for hands-on work on a shared HPC system; it is not a product, and it does not run unchanged anywhere but Delta.

What is in it

Path What it is
workspace-template/ What each participant's workspace starts from
workspace-template/AGENTS.md The house rules every agent harness reads first: how to work, what never to do, and why
workspace-template/guide/ One short script per session module (launch, orient, plan, environment, build, run, analyze, open-ended, wrap)
workspace-template/topics/ Four prepared topics (GPU training, a physics-informed network, LLM batch inference, compiled OpenMP/MPI code), each with a job script and a checker, plus the open-ended part
workspace-template/PERMISSIONS.md What the agent runs freely, what asks first, what never runs, and which files decide
workspace-template/FAQ.md, PROMPTS.md Short answers the agent starts from; prompts for going further
workspace-template/ON-YOUR-OWN.md The take-home guide: setting up an agent for your own project from an empty directory
workspace-template/opencode.json.in, .claude/ The permission maps for opencode and Claude Code
bin/, lib/ The ws-* helper scripts: ws-init (make a workspace), ws-submit (sbatch with the workshop's checks), ws-wait, ws-log, ws-check, ws-status, ws-progress and others
room/ The participants' quick card and the helpers' triage sheet
env.sh, workshop.env.example The environment a participant sources, and every operator setting with a comment
KIT-NOTES.md The kit's internal README, unedited: layout and conventions (it mentions tests and operator scripts that are not in this repository)

Start here

  • To see how the agent was instructed: workspace-template/AGENTS.md, then workspace-template/guide/01-orient.md.
  • To set up an agent for your own project: workspace-template/ON-YOUR-OWN.md. It has a starter AGENTS.md, starter permissions for opencode and Claude Code, the module-plus-venv environment pattern, and plain sbatch without the workshop's helpers.
  • To reuse a helper: the scripts in bin/ are plain bash with a shared library in lib/ws-common.sh.

What it assumes

  • Delta: Slurm, Lmod modules (pytorch-conda/2.12, llmflux), the cpu and gpuA40x4 partitions, and NCSA's hpc-gpt module (opencode with the NCSA Lumen model service).
  • A staged read-only copy with data, two small models and a container, which are not in this repository.
  • Operator settings in a workshop.env rendered from workshop.env.example: accounts, partitions, a reservation, the model list. <code> in the text stands for the workshop's allocation code, and /path/to/site-lumen-key for the site's model-service key file.

To adapt it to another cluster, expect to change the settings file, the module names in the topics' job scripts, and the Delta facts in AGENTS.md, FAQ.md and the room cards.

What is not here

The test suite, the dry-run simulator used to rehearse the session with scripted participants, and the operator scripts for staging and publishing are not in this repository.

Design choices worth knowing

  • One file of rules, with reasons. AGENTS.md is kept under 12 KB and every hard rule says why.
  • Evidence before claims. A job counts only with its KEY: value evidence line in the log and the scheduler's record; ws-check grades that, and exit 0 is not success.
  • A plan the participant approves in words before anything is submitted.
  • Permissions by the harness, not by the model. The examples in PERMISSIONS.md were tested against the real permission map.
  • No session clock. The facilitator paces the room; every module ends with what happened, two questions the participant might ask, and the next step with what to type.

License

Apache License 2.0; see LICENSE. Provided as is, without warranty.

Contact

Questions about Delta: help@ncsa.illinois.edu. Delta documentation: https://docs.ncsa.illinois.edu/systems/delta/en/latest/

About

Participant kit from the NCSA workshop session "Getting Started Using AI Agents for Research/HPC" on Delta (Oct 2026): agent rules, session guides, prepared topics and helper scripts

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages