MODULE 04 — AGENTIC AI · 16 AI Concepts

Guardrails & human-in-the-loop

The model cannot be made careful. The gap between what it asks for and what actually happens can be — and that gap is where every control lives.

The agent asks to do something

The problem it solves

Some actions cannot be taken back

EASY TO UNDO IMPOSSIBLE TO UNDO → NARROW REACH → WIDE ASK A HUMAN read a file draft a reply post to a shared doc email 4,000 customers delete the database issue one refund let it run

How it works

Four checkpoints, one human

input the model before it is read
Input gateScreen what arrives before it ever reaches the desk.
asks allowed refused an allow-list, not a plea
Action gateA list of what this agent may do — enforced in your code, not requested of the model.
asks the loop waits go no
The doorbellFor irreversible actions the loop stops and a person decides.
reply shown to you blocked checks the reply, not the request
Output gateScreens the model's own reply — catching a leaked secret or an unsafe answer before you ever see it.

So what

You are not making the model trustworthy. You are deciding, in advance and in code, which of its requests you are willing to honour.