AI agentic computing is available on the cluster using Goose and locally hosted Large Language Models (LLMs).

With Goose, users can describe a task in plain English and the agent will work with the local LLM to:

• Generate code in a programming language of your choice
• Execute the generated code
• Analyze results
• Debug and fix errors
• Iterate until the task is completed

This can significantly reduce the time required to go from an idea to a working solution, often reducing tasks that traditionally take days or weeks to only a few minutes.

Interactive Usage

• Log in to a compute node.

• Start a Goose session:

goose session

• Describe the task you would like Goose to perform.

Example:

Create a Python program that analyzes all NetCDF files in this directory, generates summary statistics, and produces plots of the key variables.

Non-Interactive Usage

Goose can also be used non-interactively and may be suitable for batch workflows.

Example:

goose run -t "Review the files in this directory and create a README.md explaining their purpose."

Reviewing AI-Generated Code

As with any automatically generated code, results should be reviewed before production use.

Users are encouraged to use the Microsoft 365 Copilot license included with Office 365 to perform an additional review of generated code for:

• Logical errors
• Performance issues
• Maintainability

One approach is to place the generated code in a OneDrive folder and ask Copilot to review the contents.

Even when code is not AI-generated, an additional review can often identify potential improvements.

Notes

Output Token Limits

Goose sessions currently have an output token limit set at 32,768 tokens.

If a task produces very large amounts of output, consider breaking it into smaller steps or multiple prompts.

Monitoring LLM Server Utilization

LLM server utilization can be monitored here

Metrics include:

• GPU utilization
• LLM token performance

These metrics can be useful when troubleshooting performance or estimating system load.

Goose provides assistance but does not replace scientific or technical validation. Users remain responsible for verifying the correctness of generated analyses, code, and conclusions.