Skip to content

Shared Context & Settings

To make agents reliable on our systems, RCAC gives them two things up front: shared context that describes the cluster, and per-harness settings that encode a sensible permission policy. This section publishes both, verbatim, as the canonical source of truth — so you can see exactly what your agent is told and how it is constrained, and so you can help improve it.

What /etc/agents.d is

/etc/agents.d/ is a directory of Markdown files on each cluster that describe that cluster to an agent: the operating environment, the filesystems, the module system, the scheduler, and the policies. Agentic tools already look for context files in well-known locations; /etc/agents.d is how RCAC supplies cluster-specific facts and prohibitions so the agent "absorbs the cluster's rules before you ask your first question."

Each file follows the same shape: a fact → the correct command or path → an explicit "do not…" → the rationale. They are written to the agent, not to you.

Canonical source of truth

The files published here are the single source of truth. They reach a running agent two ways, depending on how you work:

  • On the cluster. Cluster configuration management (Puppet) copies these files out to /etc/agents.d/ on the cluster's hosts, where they are concatenated into a single AGENTS.md and symlinked to the other well-known context filenames (CLAUDE.md, GEMINI.md, …) that each harness reads.
  • Locally. The rcac-mcp server reads /etc/agents.d/ over your SSH connection and injects the concatenated context to your agent as a read-only resource — so a locally-run agent gets the same guidance without you installing anything.

Because this repository is canonical, editing the files here is how the deployed context changes.

How to contribute feedback

This is early, evolving work, and the context and settings improve fastest with input from the people using them. If a fact is wrong, a prohibition is missing, or a setting is too strict or too loose:

  • Context Files (/etc/agents.d)

    The five context files — unix, filesystems, lmod, slurm, and policies — published verbatim (Gautschi shown as the example).

    Context Files

  • Harness Settings & Permissions

    A starting-point settings file for each of the five harnesses: allow-list the sanity commands, deny destructive operations.

    Harness Settings & Permissions

Back to Agentic AI.