04 · Practice

 Now you make the calls.

 Classify real scenarios, trace a loop by hand, stand up the free toolchain, and check yourself against interview-level questions.

 Exercise A

 Classify the scenario

 For each business scenario, pick the lowest level that genuinely solves it: rules (no LLM), a single LLM call, a workflow, or an agent. Answers are “suggested” — in the debrief, your reasoning counts more than the label.

 Exercise B

 Trace the loop by hand

 Task: “Book a room for 6 people next Tuesday afternoon.” On paper (or in your notes), write the observe → reason → act → observe trace. Mark the stop condition and where a human gate belongs. Then compare.

 Reveal an annotated trace 

 OBSERVE goal: room for 6, Tue afternoon. Rooms + calendars unknown.
REASON need candidate rooms before times → query room system
ACT rooms.search(capacity>=6) → [Maple, Peachtree, 201]
OBSERVE three candidates returned
REASON check Tuesday-afternoon availability for each
ACT rooms.availability([...], tue_pm) → Maple 2–4, 201 3–5
REASON two options; Maple earlier + has a screen → propose Maple 2pm
HUMAN GATE ⏸ "Book Maple, Tue 2–2:30?" → approve
ACT rooms.book(maple, tue_1400) → confirmed
STOP ✓ a confirmed booking exists ← the stop condition

 What graders look for: the loop actually loops (an observation changes the next action); the stop condition is named , not implied; the external action (booking) sits behind the human gate. A trace with no stop condition is a red flag.

 Hands-on

 Your toolchain — three pieces, all free

 Step 1 Gemini free API key
 Google AI Studio → create API key. No billing required; free-tier rate limits are fine for class. Keep it secret — never commit it to a repo. Already have an old Gemini key? Create a fresh one for this course.
 aistudio.google.com/apikey ↗ 

 Step 2 GitHub Copilot
 The free plan includes agent mode; verified students get Copilot Pro via the Student Developer Pack. Enable it in VS Code.
 Student Developer Pack ↗ 

 Step 3 Hello, agent
 A minimal loop — model, one tool, stop condition:

 while not done:
 thought = llm(context)
 action = parse(thought)
 result = run_tool(action)
 context += result
 done = stop_condition(result)

 Expect it to work ~most of the time. That gap between “demo” and “dependable” is not your bug — it is the field’s central open problem, and it is where this course spends Weeks 8–10.

 Before next week

 Assignment & logistics

 Do

 Read Anthropic, Building Effective AI Agents ↗ — free, no login. The workflow-vs-agent distinction and the five patterns come from here. 
 Watch Bornet, Agentic Artificial Intelligence ↗ (LinkedIn Learning) — free with GSU login, but enter through the GSU portal ↗ first and sign in with your CampusID; going to linkedin.com directly shows a paywall. 
 Sign up for Agent Radar on iCollege — student news briefings start Week 2 
 Set up your Gemini key + Copilot + run the hello-agent 
 Start thinking about your capstone — browse the Spring 2026 student project demos on iCollege for ideas, and start your one-page agent spec 

 Coming up
 Week 2 — Learning from the Best: Unpacking Coding Agents. We reverse-engineer Codex, apply the lens to Claude Code, and build a Job Search Agent three ways. Agent Radar begins (student news briefings). GA1 releases.
 Go to the Week 2 site → 

 🎯 Take it to your final project: your capstone must be a real agent — the LLM owns control flow — with a named stop condition and one human gate. Write the one-line justification now: “This must be an agent, not a workflow, because …”

 Interview check

 Could you answer these in a job interview?

 Each question is fair game in a 2026 tech or consulting interview. Try answering out loud, then reveal what a strong answer covers.

 Take-aways

 Week 1 in five lines

 An LLM predicts tokens; an agent wraps it with tools, a loop, and memory so it can act.

 The distinction that matters: who owns control flow — a developer (workflow) or the model at runtime (agent).

 Autonomy is a dial; choose the lowest level that solves the problem .

 Enterprises adopted AI everywhere but profit thinly — agents are the bet to close that gap, and reliability is the obstacle.

 Every real agent needs a stop condition , a human gate , and an honest evaluation .

 Next week: Learning from the Best — Unpacking Coding Agents. We reverse-engineer Codex , apply the same lens to Claude Code, and start building the Job Search Agent (GA1 releases).

 Week 2 readings: 

 · Codex’s real system prompt — gpt_5_2_prompt.md ↗ 

 · OpenAI Cookbook — Planning for Long-Horizon Agents ↗ 

 · Anthropic — Claude Code auto mode ↗ 

 Read them with Week 1’s question in mind — who owns control flow? — and start on the Week 2 prep page .