Acceptable Use & Etiquette¶

Running an agentic coding tool on RCAC's systems does not change the rules — it raises the stakes. An agent can issue commands faster than you can read them, so the same acceptable-use and good-citizen expectations that apply to you apply to any agent acting on your behalf. You are accountable for everything your agent does under your account, exactly as if you had typed it yourself. The rules below are the etiquette for using these tools responsibly on shared HPC resources; they are consistent with RCAC's existing Acceptable Use policy, quoted at the bottom of this page.

Login (front-end) nodes are shared by every user for editing, compiling, and submitting work — not for running it. An agent that compiles a large project, launches a parallel run, or spins up a long-lived process on a login node degrades the system for everyone. Keep agent activity on the front ends light, and push real work to the scheduler. This mirrors RCAC's Running Jobs guidance:

Important

Do NOT run large, long, multi-threaded, parallel, or CPU-intensive jobs on a front-end login host. All users share the front-end hosts, and running anything but the smallest test job will negatively impact everyone's ability to use the cluster. Always use SLURM to submit your work as a job.

Run real work through Slurm, with time limits¶

Any nontrivial computation an agent performs SHALL go through the Slurm scheduler, not run interactively on a login node. Instruct your agent to submit batch or interactive jobs (sbatch, sinteractive) with an explicit --time limit and a correct account, partition, and QOS. A bounded time limit protects both your allocation and the shared queue from a runaway job an agent started and then lost track of.

Keep sensitive data out of world-readable locations¶

Do not let an agent place sensitive or restricted data in world-readable paths (for example, a group- or world-readable scratch directory). Agents readily create scratch files, logs, and intermediate outputs; make sure those land in appropriately-permissioned locations. If your work involves regulated or export-controlled data, confirm it is handled on an approved system before an agent ever touches it.

Never expose secrets or credentials¶

Do not paste passwords, SSH private keys, API tokens, or other credentials into an agent's prompt or context, and do not point an agent at files that contain them. Whatever an agent can read, it may echo back, log, or transmit to a model provider. Treat the agent's context window as potentially observable, and keep secrets out of it.

Confirm destructive and irreversible actions¶

Require your agent to pause for your confirmation before any destructive or hard-to-undo operation — deleting files (rm), overwriting data, cancelling other users' jobs, or changing permissions on shared paths. Configure your harness so these actions prompt rather than run automatically; the per-harness settings we publish deny the most dangerous operations outright as a starting point.

Be a good citizen¶

You share these systems with thousands of other researchers. The same courtesy you extend as a human user extends to your agent: do not stress the filesystem with runaway metadata operations, do not leave orphaned processes on login nodes, and do not consume shared resources you are not actually using. If an agent's behavior is affecting the system or other users, RCAC staff may stop it without prior notice, exactly as they would any other disruptive workload.

Acceptable Purdue IT Research Resource Use¶

RCAC requires all users of its research computing resources to submit their jobs through the provided queuing systems. Do not attempt to bypass or hinder these systems. Documentation for the queuing systems on each major computing resource is available on this web site. Please be as accurate as possible about the resources which your jobs will require, as this will help the system run more efficiently and may help your jobs run more quickly. You must specify the actual number of processor cores that each job will use when submitting your jobs. Failure to do so could result in poor performance and may adversely impact other users' work.

All users of research resources must also comply with Purdue IT's Resource Acceptable Use Policy, Purdue University Policy V.4.1, as well as Purdue's Remote Access to IT Resources Policy, Purdue University Policy V.1.6.

Users of research resources should also refer to the Purdue University User Credential Standards for specific requirements on user account passwords.

Data about activity on research computing resources is routinely logged, archived and analyzed, to provide feedback about system performance, resource and software utilization, and better optimize the resources. This includes, but not limited to, environment modules loaded, applications run, hardware performance counters, disk storage consumed, and the contents of batch job scripts.

Back to Agentic AI.