Objective
Make the day's biggest engineering judgment call — should this be a plain script or an AI agent? — fast and physical. Students sort real tasks and defend each choice out loud using the trade-off: a script is cheap, predictable, and reliable but rigid; an agent is flexible but pricier, slower, and less predictable. By the end they'll instinctively reach for the simplest thing that works.
The setup
- Split into teams of 3–4. Give each team two paddles (or just two sheets): one reads 🧾 SCRIPT, one reads 🤖 AGENT.
- You (the coach) hold the Task Cards — one task per card. Have Round 1 (obvious) and Round 2 (tricky) cards in two separate piles.
- Post the trade-off where everyone can see it: Script = fixed steps, tidy input, cheap, trustworthy. Agent = open-ended, messy input, reasons on the fly, pricier & less predictable.
The rules — read aloud
Round 1 · Obvious cases
- "I read a task. On 'three… two… one… hold!' every team raises SCRIPT or AGENT — no changing after."
- "I'll point at one team: defend it in one sentence using cost, reliability, or flexibility."
- Cards: "Add two numbers" (SCRIPT) · "Convert 100 files to uppercase" (SCRIPT) · "Research 3 colleges and email me a comparison" (AGENT) · "Monitor the news and alert me when my topic trends" (AGENT).
Round 2 · Tricky / hybrid — debate it
- "Now the answer isn't clean. Raise a paddle anyway, then argue it out — there may be a right answer for either depending on assumptions."
- Cards: "Rename 500 photos by the date inside each one" (script if the date's in the metadata; agent if it must read messy handwriting) · "Sort my email into folders" (script by rules; agent if it must understand meaning) · "Summarize this one paragraph" (a one-line LLM call — barely an agent) · "Find the survival rate for women in first class" (one pandas line = script; a question-answering data agent = agent).
The process
- Round 1 (~4 min): run 4–5 obvious cards fast. Award a point for the correct paddle and a good one-sentence defense. Keep the pace snappy and loud.
- Name it: "Notice the pattern — fixed steps and tidy input went SCRIPT; open-ended or messy went AGENT."
- Round 2 (~5–6 min): play the hybrids. Let two teams disagree and each defend. Reward the team that names what assumption flips the answer.
- Land it: "The pro move isn't always reaching for AI — it's picking the cheapest tool that reliably does the job."
The debrief
- Which factor decided most Round 1 cards — cost, reliability, or flexibility? Which one broke the ties in Round 2?
- Give one task where paying more for an agent is clearly worth it. Now one where an agent would be a waste.
- Finish this: "Use the simplest thing that ______." (works · does the job reliably · is cheap and predictable)
WORKSHOP TIE-IN: This is exactly today's build choice. Your DataExplainer is an agent — it reasons about a fuzzy English question and picks a tool. But "get the average age" is a script: one line, df.groupby("Sex")["Age"].mean(). Same data, different tool — you choose based on how open-ended the question is.
WATCH-OUT: Don't let "AGENT" win just because AI sounds cooler — push teams to justify the extra cost and unpredictability. If a five-line script does it, that's the right answer, and saying so is the senior move. Keep Round 2 debates short so the energy stays high.