- Second OLS peer mentoring cohort after scholarly communication
- Agentic coding became practical recently
- Let librarians build their own tools and (potentially) improve their workflows
- Few CUNY librarians had used them
Agentic AI for Library Practice
Spring 2026 Showcase
Stephen Zweibel · Alevtina Verbovetskaya
CUNY Office of Library Services · May 18, 2026
Where this started
Have you used these?
- ChatGPT, Claude, Copilot, Gemini
- You type, it “types” back
- Predicts the next word, over and over
- Trained on text, learned patterns
- Fluent, but not grounded
Without tools, the model can only estimate from training data.
Reaching for groundedness
- Searches the web
- Reads files you upload
- Runs code
A tool is a function the system can call. “Search this,” “fetch that.” That's tool use.
Same question. This time it actually looks it up.
Agentic means it loops
- Regular: ask, answer (with maybe one tool call)
- Agentic: a harness loops. Runs the tool, feeds the result back, prompts again
- Stops when done, or stuck
- Harnesses: Claude Code, Codex, Cursor
What that looks like
The librarian asks: “Find recent books on music theory we don't already own.”
- The tool searches WorldCat for recent music theory titles
- It checks Primo by ISBN to filter out titles already held
- It fetches publisher links
- It returns a structured list (title, author, publisher, ISBN, link)
The librarian reviews the list and decides what to order.
The program
- Sixteen weeks
- Weekly meetings, plus office hours
- Librarians across CUNY
- Mixed roles, mixed backgrounds
- Each brought a workflow from their job
- Curriculum reorganized around what they hit
- Tools paid out of pocket
What we wanted from it
- Librarians who can build with these tools and judge what they produce
- People across CUNY campuses who can answer agentic AI questions for colleagues
- Practitioners who decide how agentic AI fits into their work
Use cases
- Applications
- Workflows
Learning by doing
- Experimentation
- A way to understand what these tools can and can't do
- Helpful even if you're not building applications
The OER and what's next
- Working with AI Coding Tools: Guides for Librarians and Researchers · CC BY 4.0
- Seventeen guides, including:
- Talking to the tool, recognizing failure patterns
- APIs, scraping, tabular data
- CLAUDE.md, Skills, verifiable goals
- Article in draft
- What librarians do next, as these tools do more
Today's presenters
- Shamiana Pond · OLS · A Decade of ACRL
- Mark Eaton · Kingsborough CC · Recycling the Old into the New: Using Agentic Coding to Extend RSS
- Kristen Fredericksen · Central Office · From Manual Process to Web App: Automating CUNY's Shared Print Retention Transfers
- Tokunbo Adeshina · Bronx CC · Enhancing Library Workflows: Claude Code AI Use Case
- Anthony Bishop · BMCC · Reference AI Agent
- Mariya Gluzman · Brooklyn College · Human-in-the-Loop Accessibility Remediation Tool
- Patty Sumire McGowan · BMCC · More Work to Do
- Robin Naughton · Queens College · Augmenting Website Workflows
- Karen Okamoto · John Jay · Using Claude to Quickly Create a LibGuides Widget
Thank you
to Allie Verbovetskaya, Michael Waldman, Maura Smale, and the cohort