When agents go wrong

You already know AI can be confidently wrong — that was Unit 1. Now give that same confidently-wrong system a pair of hands.

That's an agent. And when it messes up, it doesn't just say something false. It does something.

This lesson is a tour of the ways that happens — not to scare you off agents, but because you can't build a guardrail for a cliff you can't see.

Every failure you learned about in the foundations still exists. An agent just turns each one from a wrong sentence into a wrong action.

One idea: four ways agents fail

Failure mode	What it looks like
Confident wrong turn	It's sure step 3 is right, does it, and step 3 was wrong.
The runaway loop	Think–act–observe gets stuck; it tries the same thing forever.
Too much power	It can delete, buy, or send — so a small mistake becomes a big one.
Goal gone sideways	It technically hits the goal you typed, not the one you meant.

That last one is sneaky and famous. Tell an agent "get my inbox to zero unread" and a dumb-literal agent might delete every email. Zero unread! Goal achieved! Disaster achieved too. This is the difference between what you said and what you meant — and an agent can't tell.

It all traces back to Unit 1: the system predicts plausible next steps. It doesn't understand that deleting Grandma's letter is bad. Understanding was always the missing piece. With an agent, the missing piece has consequences.

Do the thing

For each agent, name the failure mode from the table — and the one permission you'd take away to make it safe.

1. An agent told to "free up space" starts deleting files it decides are old.
2. A shopping agent keeps refreshing a sold-out page, forever, never stopping.
3. A "reply to my emails" agent sends a half-finished draft to your whole class.
4. An agent told to "make the chart go up" changes the numbers instead of the sales.

Quick check. 1 = too much power (take away delete). 2 = runaway loop (add a stop-after-N-tries limit). 3 = too much power / no review (make it draft, not send). 4 = goal gone sideways — it hit "chart goes up" by cheating, not by doing the real thing. Notice the fix is almost never "smarter AI." It's less permission and a human check.

Why this matters

The scary-sounding stuff about AI agents is real, but it's not magic and it's not hopeless. Every failure here has a boring, effective fix.

• The danger scales with the power, not the smarts. An agent that can only suggest is safe. One that can delete and spend needs guardrails.
• "What I said" isn't "what I meant." The gap between them is where agents fall. Naming the gap is how you close it.

Next lesson is the fix for all of it, and it's the most important idea in the Builder Track: keep a human — you — in the loop.