MODULE 04 — AGENTIC AI · 17 AI Concepts

Where agents actually work

Five jobs that look unrelated. Pick any of them and the same ring from page 13 appears underneath, with different words in the same five slots.

Pick a job

What the good ones share

Three properties, or don’t bother

tests pass tests fail no judgement needed the loop can grade itself
A checkable resultSomething outside the model can say whether it worked.
3 failures, then one works and nobody was harmed retry is free
Cheap, reversible attemptsFailure costs another go, not an apology to a customer.
6/10 at 10 it stops and hands over never “until solved”
A hard stopping ruleA budget it cannot argue with, because judging “done” is its weakest skill.

Not every AI use is a loop

Two shapes, not one

Everything above runs the ring until a check says stop. Plenty of AI use never loops at all — one prompt, one answer, done.

runs until a check says stop
The ringCoding agent, support triage, research, monitoring, inbox — every job on this page so far.
one prompt, one answer, done
One straight lineTurning a prompt into a picture, translating a sentence, summarising a page — no check, nothing to retry.

So what

A loop is for jobs with no fixed endpoint. A single call is for jobs that already have one — knowing which one you're building is the actual design decision.

MODULE RECAP

Before you move on

Click an idea below, then click the card it matches.

An agent is just this loop, with memory, teammates, and a leash — running on jobs you can actually name.

Continue reading →