Module 1 · Foundations of Agentic AI · scripted
Exercise: Your First Agents
What "agency" means
Agency is the capacity to decide and take the next action yourself, instead of waiting to be told what to do. An LLM is showing agency when it drives the conversation — deciding what to ask, what to do next, and when it is done. That capacity is what the name agent points at: an agent is something that acts on its own behalf.
Agency, in one line
An LLM with agency drives the conversation. That is what the name "agent" points at.
The goal: prompts that produce agency
In this exercise you will practice building an agent by writing flipped interaction prompts. The goal is to have the agent drive the conversation — and the moment it begins driving, it is demonstrating agency. You are practicing writing and architecting conversations that lead to agency from the agent.
What you are practicing
An agent here is a prompt — nothing more. Your prompt must make the LLM take charge: decide what to ask, issue one action at a time, wait for the report, and choose the next action based on what came back. When that happens, the conversation has a driver — and it isn't you.
Where to run it
Use either environment:
- An LLM of your choice (ChatGPT, Claude, …) — paste your prompt as the first message of a fresh conversation.
- The Agent Lab — create an agent, paste your prompt as its system prompt, and converse. Agents and their conversations are saved, so you can build several and switch between them.
Where to run it
Scan to open the Agent Lab
If you use the Agent Lab, it's worth toggling between the Chat view and the Prompts view as you go. The Prompts side shows, for every Send, the entire prompt that was actually transmitted — Prompt 1, Prompt 2, Prompt 3… — system message and every turn included. Watch it as your conversation proceeds: nothing is happening except that one growing prompt being re-sent, numbered, again and again.
One thing worth experimenting with as you go: the driving prompt can live in different places — in the system instructions, as the first user message, or even dropped in as a user message mid-conversation ("From now on: guide me through X, one step at a time…"), which can flip an ordinary conversation into a driven one on the spot. If you try more than one placement, compare: does it change how strongly the agent drives, how well it holds the pattern over many turns, how easily it slips?
Step 1 · The cooking guide
Step 1 · What you'll do
Write a flipped interaction prompt whose job is to guide you through cooking a dish, one step at a time. The basic pattern:
The flipped interaction pattern
Guide me through cooking ___ (or: a dish of your choosing).
That skeleton is a starting point — rewrite it in your own words and sharpen it. Then run it, playing the cook: report honestly what would happen at each step, and at least once report that something went wrong ("the pan is smoking" · "there's no butter"). The test of your prompt is whether the agent keeps driving: gathers what it needs, issues actions, adapts.
Now run it
Rewrite the skeleton in your own words, then see if you can get it to actually teach you to cook the dish. Report honestly what happens — including, at least once, a problem ("the pan is smoking" · "there's no butter"). The test: does the agent keep driving — gathering what it needs, issuing actions, adapting?
Capture — end of Step 1
Step 2 · Add a format
Step 2 · What you'll do
Now modify your prompt to add a FIXED FORMAT that the agent must output every step in.
Now put structure on the agent's output. Extend your prompt so every turn from the agent follows a fixed format — for example:
A step format
Run it again. Notice what the format buys: every step now arrives with an explicit request for the feedback the agent needs — the step and its report are designed as a pair.
Capture — end of Step 2
- Was the agent able to hold and maintain the format?
- If it got the format right in the first few steps, did it keep it all the way through?
- If it got it wrong early in the conversation, did the errors compound from there?
Step 3 · A complex process from your domain
Step 3 · What you'll do
Now, for your domain: think of an interesting process the agent could guide somebody through, step by step — and a format for outputting the information. Build a prompt that guides them through it.
The opening interview (questions before the guidance starts) is optional this time — include it only if your process needs it.
Repeat the design — not cooking. Choose a complex process you know well: calibrating an instrument, running an analysis choice by choice, preparing a submission, a lab protocol, setting up field equipment. Write a flipped interaction prompt that guides a person through it step by step, with your output format, and adapts to what they report. The opening interview is optional this time — include it only if your process needs it.
Run it as the person being guided. Give it one honest run — and one report of a problem the standard documentation for that process never mentions.
Capture — end of Step 3
Deliverable
At the end of the exercise, each pair has three captures — the three prompts plus their evidence. We will read the prompts together and discuss them at the end: what made the strongest prompts drive, where agents slipped out of the pattern, and what one sentence most changed an agent's behavior.
- Where in the conversation did agency actually appear — which turn, and what in your prompt caused it?
- If you tried more than one placement (system instructions, first message, mid-conversation), what differed?