Agentic AI at RCAC¶
Researchers on RCAC's clusters use agentic coding tools (Claude Code, OpenAI Codex, Gemini CLI, opencode and Warp) to write Slurm scripts, debug failing jobs, move data and explore software. RCAC does not prohibit these tools. It publishes cluster-specific context that makes their output more likely to be correct on RCAC systems, and guidance on checking that output.
These tools can produce answers that look correct and are wrong: a Slurm script for a scheduler RCAC does not run, a module name that doesn't exist, a path on the wrong filesystem. This section describes the context RCAC provides and how to check what an agent produces. Review what the agent proposes, and ask it to explain its choices.
New and changing
The MCP servers and the shared context files are new and will change. See the shared-context pages for how to send corrections.
In this section¶
-
Acceptable Use & Etiquette
The rules for running agents on RCAC systems, consistent with Purdue's Acceptable Use Policy.
-
Best Practices & Limitations
Using an agent for research versus operations; giving it context; checking output; and the risks of letting an agent run commands.
-
MCP Servers
RCAC's MCP servers (
rcac-mcp,globus-mcp,rcac-docs-mcp) and what they let an agent do. -
Running Agents
Set up your harness on the cluster (login nodes) or on your own machine, reaching the cluster over SSH, for all five harnesses.
-
Shared Context
The context files RCAC places on each cluster, and how to connect your harness to them.
For cluster-specific setup, see the Using AI Agents chapter in each cluster's user guide: Anvil, Gautschi, Gilbreth, Negishi and Scholar.
Reference: Lentner and Ashish, "Hello Computer: HPC in the Agentic Era", PEARC'26 (2026); replication package.