Student Intern — AI Agents for Researchers

Researchers from all over the world use SDSC’s supercomputers, and most of them are experts in their own field rather than in schedulers, module systems and job-script syntax. A wrong resource request or the wrong queue doesn’t just fail their job. It can clog a shared queue for everyone else.

I’m building agents that sit on the command line and handle that part for them: requesting compute nodes without needing to know every parameter, loading environment modules without untangling dependency chains, and avoiding mistakes that waste shared resources. The work is in hpcgpt-cli.

It’s half engineering and half research. Alongside building the tools, I’m measuring token usage, task success, and what changes when the agents get more system privileges. The open question my supervisor and I keep returning to is how much safety you give up by letting an agent write to the system instead of only reading from it. It’s the same question VibeAuthBench asks, from the operator’s side instead of the developer’s.

The work is ongoing and still early, so there’s more to come here.