The 42nd Period | STEM Education Hub

Quick reference for the whole activity: chatbot, automation or agent, the loop, the permission ladder, the log, how errors compound, and the layers of guardrails.

AI That Acts infographic: chatbot versus fixed automation versus agent, the goal plan act observe decide loop, the permission ladder from read to hard to undo, ask first, the log, errors that compound, and layers of guardrails

🧭 Plot Summary

A chatbot writes. An agent does: it can book the room, send the email, delete the file or buy the tickets. That is useful, and it changes what a mistake costs. A wrong answer on a screen is one thing. A wrong email to 600 people is another.

You'll learn how an agent works, then run an Agent Sandbox: an invented agent with a job, a hidden instruction it might follow, and a typo it will make. You set its permissions and guardrails, run it, and read its log. The sandbox has no AI inside. A script plays the agent so the effect of each setting is clear.

What you'll do in this activity

Tell a chatbot, a fixed automation and an agent apart, and name the parts of an agent.
Sort actions by how hard they are to undo, and set permissions: Off, Ask first or Auto.
Run the Agent Sandbox: an invented agent, a hidden instruction and a built-in mistake. Change the settings and record what happens.
Read an agent's log to find the first wrong step and decide whether it could be undone.
Write a guardrail plan with layers, review points and a way for each layer to fail.
Add an AI-use statement that says what you checked and which decisions were yours.

Ground rules

Use the Agent Sandbox. Don't connect any real agent to your real email, files, accounts or money. Use an approved chatbot only to plan in text.
Don't type passwords, personal details or anyone else's private information into a chatbot.
When an agent asks you to approve something, read it first. An approval is a decision, not a button.
Everything in the Sandbox is invented. Real tools differ, so ask what a real tool can do before you trust it.

Why it matters

More tools now act for you, and many are called "agents." Knowing what to ask, such as what it can do without asking, what can't be undone, and what gets logged, lets you decide how much to trust one.

✅ Self-Check Before You Share

Check off each item as you get there. These aren't grades. They're your own signal.

I can say how an agent differs from a chatbot and from a fixed automation: it uses tools and chooses the next step, in a loop.I can name the six parts: goal, tools, permissions, context, log and stop rule.I sorted actions into read, draft, change I can undo, and hard to undo, and I said which should be Ask first or Off.In the Sandbox I set every tool to Off, Ask first or Auto, and I wrote why.I ran the Sandbox at least three times and recorded what changed each time.I can read an agent's log, find the first wrong step, and say whether it could be undone.My guardrail plan stacks layers, and for each layer I wrote how it can fail.My AI-use statement says which tool, what for, what I checked, and which decisions were mine.

🧪 Graded Assessments

🛡️ Agent Plan (PDF)
✅ Agent Check

🧠 Instructor Insight

📄 Setup Notes (PDF)
▶️ YouTube Video
🎥 Instructor Video

🤖 AI Study Tools

For review and clarification. Support your understanding, don't replace the work.

📺 AI Video
🎧 AI Podcast
🧩 Blooket Review

📜 Field Note

Before you give an agent a job, ask: what can it do without asking me, and what can't be undone? If the answer is "anything" and "a lot," it needs more limits.

🔬 Deep Dive

Ready to go further? The Deep Dive has a Loop lab, a Ladder lab, a Compounding lab, and the Agent Sandbox where you set permissions and guardrails and read the log.

Open Deep Dive →

📝 Formative Activity

A 15 to 20 minute check: sort chatbot, automation and agent, climb the ladder, read an invented log, and pick the guardrail that fits.

⏱ Start the Check →