Published: 2026-06-24
Analysis & perspective

How a Former NYU Professor Built a 34-Agent Team With Claude Code

Chapters / key moments (click to jump — plays here on the page)

In this conversation, AI advisor Allie K. Miller and former NYU data-science professor Dr. Andrea Jones-Roy discuss how Jones-Roy went from "just using ChatGPT" to running a 34-agent "digital workforce" in Claude Code — wired to Gmail and Google Workspace through MCP. It's a candid look at agentic onboarding, meta-prompting, and keeping critical thinking in the loop, rather than a step-by-step build.

Source video

"How a Former NYU Professor Uses Claude Code" by Allie K. Miller — Watch on YouTube →

Key Takeaways

She built 34 agents by first building an agent-builder. Rather than hand-writing each one, she created one meta-agent (using Claude's /skill creator) that then generated the prompts for all 34 — "my wish from a genie is more wishes."
A research agent emails her a daily/weekly digest. It surfaces new AI developments in her field — political science and AI-for-science — by pulling from Gmail and Google Workspace through MCP connectors she set up.
Claude Code runs on her own machine as the agent harness. She confirms she chose Claude Code (over Codex), brought it onto her computer, and connected it to tools via MCP — then dictates ~90% of her own text and lets agents draft.
She reframed AI coding from "cheating" to liberating. Her "soul" is in the data analysis and the architecture, not in hammering out Python syntax — so delegating the code felt like "motorcycling down the coast" instead of biking uphill on a rusty bike.
Her rule for responsible use: lead with your brain. Be able to explain every decision (why this data, why this analysis), keep critical thinking in the loop, and never lean on AI under deadline pressure — "urgency is poison."

New to wiring agents up to your tools? See our skills guide, what MCP is, and real-world agent setups for the hands-on version.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Build the agent builder first
The guest built a team of 34 agents in Claude Code by first creating one agent whose job was to write the prompts for all the others, rather than hand-writing each.
Use skill-creator for meta-prompting
She uses the /skill-creator skill inside Claude as the builder of other skills.
Start from a real need
Her first useful agents came from one need: keeping up with research on how AI is used in politics and science. They send her a plain email with the day's or week's updates; the host points out this means Claude Code is connected to Gmail, which is done through MCP.
Use it for reminders
She also gets AI-driven pings for things that aren't on her calendar, such as who she needs to reply to.

Gotchas

She does not use AI to write her comedy material and feels uneasy presenting agent-generated analysis as her own; she uses it freely for coding.
This is an interview; the setup isn't shown step by step, so treat the pattern (builder agent, skill-creator, email digests) as the takeaway.