Unit 6Looping and agents

The big idea

Looping means having AI check its work against your standards and improve it, over and over, until it's good. I call it prompt engineering on steroids.

Why it matters

This is the difference between using AI and directing it. Someone who can run a simple plan-build-review loop, even with two free chat windows, gets far better work than someone who takes the first answer.

The loop: four roles

1. The boss makes the plan and sets the standard. Think of a recipe. The boss writes it.
2. The worker does the work, following the recipe.
3. The reviewer checks the work against the standard, like comments on a report card.
4. The fixer applies the feedback and sends the result back to the boss.

Then it goes around again. About three rounds is usually enough. More rounds cost more time and money and add less each time.

Use different AIs for different roles. They think differently, so one catches what the other misses. It's like getting a friend to check your work instead of checking it yourself.

You own the start and the end. You write the plan and the standard at the start. You review the final result at the end. The middle can be AI.

Do this: the two-window loop

You don't need any special setup. You need two chat windows, ideally from two different AI companies.

1. Window 1 (boss): "Here's my goal: [goal]. Write a step-by-step plan and the standard the result must meet. Don't do the work."
2. Window 2 (worker): Paste the plan. "Do step 1 of this plan."
3. Window 1 (reviewer): Paste the result. "Review this against the standard. List what's wrong, most important first."
4. Window 2 (fixer): Paste the feedback. "Fix these issues."
5. Repeat steps 3 and 4 up to three times, then move to the next step of the plan.
6. You review the final version. Does it actually meet your goal?

What an agent is

A chatbot answers your question and waits. An agent takes a big goal, breaks it into tasks, and works through them on its own.

Think of a coffee order. "Make me a drink" gets you something random. "Medium, oat milk, one shot, no sugar" gets you exactly what you wanted. A clear task gets a clear result. An agent is just a lot of clear tasks in a row.

When I ran a live research job for a student, the system split it into separate agents: one looked at patents and intellectual property, one at legal liability, one at regulations, one at evidence from real pilots. Each one had a clear job.

What agents are made of. An agent is basically a text file, like a Build-A-Bear. It says who the agent is, how it works, what it needs, what it checks, and what it produces. A "skill" calls several agents in order to get a result. All AI is files. The model just gives them the power to do things.

Questions to ask yourself

• Where in my project am I accepting the first answer?
• Is my reviewer actually different from my worker, or am I checking my own work?
• How would I break my project into tasks a stranger could do?

In my words

"If you just go ahead and you're like, I'm going to do the bare minimum, AI's going to do all my work for me, you're going to have very average results."

"Break it down. Don't try to do it all at once."

Worksheet

Use Worksheet 6: Two-window loop tracker in the PDF kit.