L

Learn AI

Track progress · learn offline

Open

Lesson 6 of 7

Keeping agents on a leash

7 min read

An agent that can book a table can also, in theory, empty your bank account. So how do you get the help without handing over the keys?

Let it read; gate what it can break

The fix is a leash: clear limits on what the agent may do on its own. Low-risk actions — searching, reading your calendar, drafting a reply — can run freely. But anything that spends money, sends a message as you, or can't be undone stops at a checkpoint and waits for your OK. This is human-in-the-loop: the agent proposes, you approve.

Human-in-the-loop means an agent never takes a risky, irreversible action alone — a person signs off first.

Permissions and guardrails

Two more tools ride alongside the checkpoint. Permissions decide what a tool is even allowed to touch — read-only versus read-and-send. Guardrails are rules that block clearly bad moves before they run. Together they mean an agent works within a fence you set, not the whole open field.

The goal isn't a fully autonomous agent — it's a useful one you can trust, because the risky edges are fenced and you hold the final yes.

Be most careful with actions that move money, send messages, or delete things — the ones you can't take back. Read what the agent proposes before you approve; a fast yes is how automation goes wrong.

The shape of it

Your agent lines up four steps; three are read-only and one is "pay the €900 invoice." What should happen?

Continue in the app

Take the whole Agents & Tools course — tracked

Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.