About this Event
Claude Code on Midway3: An AI Coding Agent as Research Infrastructure
Abstract: Research computing is being reshaped less by language models themselves than by the agents built on them: systems that take an instruction in plain language, act on a real machine, check the result, and keep working until the task is done. Claude Code is one such agent, built to run in the terminal rather than in a browser, so it can sit where research already happens: on a shared cluster, next to the data, the software, and the job scheduler. This workshop introduces Claude Code as everyday research infrastructure rather than as a novelty. We consider what an agent is and how Claude Code differs from a chat assistant, where it genuinely helps a research workflow and where it does not, and what it means to supervise it responsibly on hardware shared with others, including how much autonomy to grant it and how to keep its work reviewable. Participants then spend most of the session working with Claude Code on Midway3: asking it about the cluster in ordinary English, using it to carry out a small data analysis, and finally running that analysis as a real batch job. Throughout, the emphasis is on judgment: knowing what to hand Claude Code, how to check its work, and how to direct it when a first pass needs correcting. Attendees leave with Claude Code working on the cluster and a realistic sense of what to delegate to it.
Learning Objectives:
- Describe what distinguishes an AI agent from a chat assistant, and judge which research tasks are worth delegating to Claude Code.
- Use Claude Code in plain language to explore and operate a shared HPC environment, reading the commands it runs rather than trusting its summary.
- Review Claude Code’s analytical work and direct it to revise or verify its own results.
- Take Claude Code-assisted work from an interactive session through to a reproducible batch job, and diagnose the jobit when it fails.
- Set appropriate limits on an agent’s autonomy when working on shared research hardware.
Presenter: Youzhi Yu
Duration: 2 hours
Level: Introductory
Prerequisites: An RCC account and basic comfort at the command line.
Event venue & nearby stays
John Crerar Library - Kathleen A. Zar Room, 5730 South Ellis Avenue, Chicago, United States
Concerts, fests, parties, meetups — all the happenings, one place.