Unit 7 · Lesson 3
Keep a human in the loop
Design the guardrails that make an agent safe to actually use.
You've seen what agents can do and how they go wrong. Here's the good news: making one safe isn't about building a smarter AI. It's about deciding, up front, where you stay in charge.
That's called keeping a human in the loop — and "the human" is you. It's the same move from Unit 5, where you were the verify step in your workflow. Now we make it a rule.
The safest agent isn't the one that never makes mistakes. It's the one that can't make a big mistake without a human saying yes.
One idea: four guardrails
You don't need all four every time. You need the ones that match how much damage a wrong step could do.
| Guardrail | What it does |
|---|---|
| Ask first | The agent must get your yes before a risky action (spend, send, delete). |
| Set limits | Hard caps: "never spend over $10," "stop after 5 tries." |
| Stay reversible | Prefer actions you can undo. Draft, don't send. Move to trash, don't erase. |
| Show the work | The agent explains what it's about to do before it does it. |
The golden rule underneath all four: the bigger and less undoable the action, the more human checking it needs. Reading your calendar? Let it run. Emptying your bank account? That needs a very awake human.
Do the thing
Design the guardrails for a real agent: "an agent that helps me manage my group project." For each power it might have, pick a guardrail:
- It can read the project docs → ______
- It can draft messages to teammates → ______
- It can send those messages → ______
- It can delete files it thinks are duplicates → ______
Quick check. Reading docs is low-risk — "let it run" is fine. Drafting is safe if it stays a draft (reversible). Sending should be "ask first" (it speaks as you, to real people). Deleting should be "ask first" and "stay reversible" (trash, not erase) — maybe both, because a wrong delete is the hardest to undo. The pattern: guardrails get stricter as actions get bigger and harder to take back.
Why this matters
You finished Unit 7. You can now think about agents the way professionals do — not "is it smart enough?" but "what is it allowed to do without me?"
- Design the permissions, not just the prompt. Deciding what an agent can't do is the real safety work.
- Reversible and 'ask first' beat 'trust me.' Hope is not a guardrail.
You've got agents. Next unit flips the camera around: instead of an agent doing tasks, you'll teach a model your own taste — using nothing but a handful of well-chosen examples.