Objective
Let students physically run an AI agent's loop — reason, use a tool, read the result, repeat — so today's theory (Planning → Tools → Memory → Action, the ReAct framework) is something they've done, not just heard. They'll feel why a chatbot with no tools guesses, while an agent with tools finds out.
The setup
- Split into teams of 4. Assign roles: Brain (says the Thought + next Action), Hands (uses a tool), Memory (writes each Observation on paper), Narrator (reads the loop aloud so the room can follow).
- Give each team the two Tool Cards face-down: a 🔎 SEARCH card (an index card with 4–5 facts you wrote) and a 🧮 CALCULATOR card (they may use a phone calc only when this tool is "called").
- Pick a Goal question that needs both tools, e.g. "How many days until our showcase, and what's one recent fact about AI in schools?" Put the answer facts on the SEARCH card.
The rules — read aloud
Round 1 · "You are a chatbot"
- "Brain, answer the Goal question in 10 seconds — from your head only. No tools, no phones."
- "Whatever you say, that's final. Memory, write it down."
Round 2 · "Now you're an agent"
- "Same question — but now you loop. Brain says a Thought then an Action: 'I should SEARCH for the showcase date.'"
- "Hands flips the matching Tool Card and reads the result. Memory writes that Observation on paper."
- "Repeat: Thought → Action → Observation. Stop only when you can answer for real. No guessing allowed — if you don't know it, call a tool."
The process
- Round 1 (~2 min): every Brain blurts an answer. Collect them. Most will be vague or wrong on the exact date/fact — that's a hallucination.
- Name it: "No tools, no memory — just a confident guess. That's a plain chatbot."
- Round 2 (~5–6 min): teams run the loop out loud. Narrators announce each "Thought… Action… Observation." Cheer the first team that reaches the goal with evidence.
- Compare: put a Round-1 guess next to a Round-2 answer on the board. The difference is the whole point of building agents.
The debrief
- What changed once you were allowed to use tools? Did you still need to "know" the answer up front?
- Which role was Memory doing the job of in a real agent? (Keeping past observations so the loop doesn't forget.)
- Finish this: "A chatbot answers in one shot. An agent ______." (loops · reasons · acts · checks · repeats until done)
WORKSHOP TIE-IN: You just ran ReAct by hand — Reason + Act. In today's Colab build, your "DataFetcher" agent does the exact same loop in code: the LLM writes a Thought, calls a tool (search_web(), summarize_text()), reads the result, and loops until the task is done.
WATCH-OUT: In Round 2, don't let a Brain sneak in a remembered answer — force every fact through a tool call. That discipline ("if unsure, use a tool") is exactly what stops real agents from hallucinating. Keep the pace fast; narrate loudly.