Supervising an Agent
Week 7
Competencies: Skilling and Productivity, Critical Analysis, Technical Understanding
Overview
You’ll give an AI coding agent a small, LIS-flavored task and watch it work.
This is a lab to do honestly. A clean output that came easy teaches you less than a messy output you understood as it happened.
Pick Your Path
In class, we’re using OpenAI’s Codex, since there’s a free tier (2025). But you can use other coding agents, such as Claude Cowork or Google’s Antigravity.
Pick a Task
Choose one of the five. Download the starter materials, unzip, and point your agent at the folder.
Messy Bibliography — ~100 citations with inconsistent author formatting, missing years, near-duplicates. Clean it into a consistent bibliography.
Metadata Crosswalk — 10 MARC records. Produce a clean Dublin Core conversion plus one-sentence notes per record on what got lost in translation.
Reader’s Advisory Bot — Build a small tool for browsing book award winners. (No starter data — sourcing it is part of the task.)
Research / Study Guide Builder — From a folder of class materials, produce a structured LibGuide-style document: background, key sources annotated, further reading. Materials are in Canvas (if that link breaks, look in Files on the course Canvas page).
Interview Simulator — Set up a “bot interviewer” based on a job posting. Collect more detail about the workplace than the posting alone gives. Iterate on the interview voice; have it save your answers and suggestions.
partway through, tou may also bring your own task — something from your work, a side project, a real folder of files.
Walkthrough
1. Set up
- Make a folder on your computer with your starter files.
- In your agent of choice, add the folder as a project (Codex: “Add New Project”; Cowork & Antigravity have the same affordance).
- Ask the agent what’s in the folder. Just look around.
- Tell the agent to initialize the project — in Codex/Antigravity it’ll write
AGENTS.md; in Cowork it’sCLAUDE.md. Read that file. What did the agent decide about your project? Anything wrong?
2. Plan first, then act
- Ask the agent to plan the task — not do it. Example: “Don’t make changes yet. Tell me how you’d approach [task] with what’s in this folder.”
- Read the plan. Is anything missing? Anything that worries you? Push back if so.
- Once it looks right, tell it to go.
3. Watch the loop
- As the agent works, watch for the moment it re-reads its own output and changes course. That’s the loop.
- Note one specific moment: what did it just do, what did it notice, what did it change?
- If it never self-corrects — note that too. Some tasks don’t need it; others do and the agent missed it.
4. Try to break it
- With a working result in hand, re-prompt with something deliberately vague, contradictory, or out of scope.
- Watch what happens: does it ask, or does it guess?
- Note one place where a vague prompt produced a confident-but-wrong answer.
5. Iterate
- Got a result? Is it good, bad, workable?
- Consider next steps — or swap to a project of your own choosing and see how far you get.
6. Capture as you go
Keep a doc open alongside the agent. Jot down, while you work:
- The plan the agent proposed (copy/paste)
- The one loop moment you noticed
- The one failure when you tried to break it
- A screenshot or two of anything that surprised you
The reflection writes itself if you capture as you go.
Completion Details
Prepare a reflection on the following questions - 1-2 paragraphs. We’ll discuss this in class, so you can have basic notes if you don’t intend to include in the portfolio.
- Which task you picked and which agent you used
- The plan the agent proposed (verbatim)
- The one loop moment you observed
- One thing that surprised you and one thing it got wrong
- A trust question — what would you need to see before you’d let this agent work without you watching?
Portfolio Details
Size: 2 points
If including in your portfolio, submit:
- The above five items as a 1–2 paragraph reflection
- The finished artifact (an output file from the task, or a zipped copy of the working folder)
- Optional: screenshots of moments that mattered
Grading focuses on the reflection — depth of observation about how the agent worked, not the polish of the output - and the quality of the final artifact.