Your First Agentic Workflow on the Cluster

24.2. Your First Agentic Workflow on the Cluster#

This walkthrough runs one small agentic task end to end, so you can see the whole loop before trusting an agent with real work. It uses Claude Code as the example, but the shape is the same for any terminal agent.

        flowchart LR
    A["Start a session<br/>and launch the agent"] --> B["Point it at<br/>your work"]
    B --> C["Give it one<br/>scoped task"]
    C --> D["Review<br/>before it acts"]
    D --> E["Verify<br/>the result"]
    E -->|refine| C
    classDef s fill:#14154C,color:#ffffff,stroke:#3D3E82;
    classDef v fill:#A51C30,color:#ffffff,stroke:#A51C30;
    class A,B,C,D s;
    class E v;

24.2.1. Start a session and launch the agent#

Before your first run, install and sign in to the agent, and read how to size an allocation, in Using Agentic AI on the Cluster.

Start a tmux session on the login node, so a dropped connection does not end your work. Inside it, start an interactive job on a compute node, set up a practice folder, and launch the agent. A small CPU allocation is enough. When Claude Code first starts in a folder, it asks whether you trust it. Trust only folders whose agent configuration you know, such as your own projects; see Untrusted repositories.

hostname                          # note the login node, to come back to it
tmux new -s agent
salloc --partition=test --time=0-02:00 --mem=16G --cpus-per-task=4
mkdir -p ~/agent_practice && cd ~/agent_practice
git init                          # git shows what the agent changed
module load python
python -m venv .venv && source .venv/bin/activate
pip install pandas matplotlib     # the packages the practice task needs
claude --permission-mode default  # manual mode: asks before it changes anything

Without --permission-mode default, new sessions start in auto mode where it is available, and a classifier approves routine actions for you. For a first run, manual mode lets you see each request; see Permission modes.

Tip

If your connection drops, SSH back to the same login node and run tmux attach -t agent. If the job has ended, start a new one with salloc and cd to the same folder. Load the environment again (module load python, then source .venv/bin/activate), and run claude --continue --permission-mode default to pick the conversation back up in manual mode; see Keeping a session alive.

24.2.2. Point the agent at your work#

The agent works in the folder you start it in. It reads files without asking, and what it reads goes to the provider, so this first run uses a new folder and made-up data. In your own projects, commit your work before you start the agent, so you can undo its changes; see Git as the undo layer. An instructions file such as CLAUDE.md tells the agent your conventions; see Configuring Agents for Your Project.

24.2.3. Give it one scoped task#

Start with a single, checkable task rather than a whole project:

Create data/measurements.csv with 200 rows of made-up hourly temperature readings, print summary statistics, and save a histogram of the temperature column to figures/temperature_hist.png.

The agent’s commands use the environment you activated before starting it. A narrow task is easy to review and to verify, and it shows you how the agent behaves before you give it anything larger.

24.2.4. Review before it acts#

In manual mode, the agent asks before it edits files or runs commands that change things. Read each request before you approve it, especially anything that deletes files, moves data, or installs software.

Stop. Press Esc to stop the agent at any point.
Approve. Choose Yes to approve one action. For a shell command, Yes, and don’t ask again saves a rule that allows it from then on in this project, so avoid it while you learn.

This review is also your defense against untrusted content; see Agent security.

24.2.5. Verify the result#

Check the output yourself: open the figure, read the numbers, and run any tests. Look for small departures from the prompt, for example a column named temperature_c when the prompt asked for temperature. An agent’s result is a lead to confirm, not a finding to trust. When you want to measure quality more systematically, see Evaluating and Monitoring Agents.

Once the task is right, refine it or move on to the next one. When you are done, type /exit to leave Claude Code, then exit to end the interactive job and release its resources, and exit once more to close tmux.