Facilitator Run-Sheets — Level 1.5 Agents Workshop

Facilitator Run-Sheet · Level 1.5 · Session 1 of 3

Meet the Agent

≈ 80 min Stations 1–3 · Deck: session-1 (press S) · Notebook: Session 1 · Energizer: Agent or Not?

Clock	Min	Block	Say / do — key cues
0:00	3	Welcome	"You've prompted AI; today you make it act." Everyone into Colab + Coach HQ.
0:03	8	Idea 1Chatbot vs agent	Talks vs acts · agent = LLM + tools + loop. The magic ingredient is tools.
0:11	6	Idea 2The engine	Quick LLM recap: tokens, context window, temperature. Don't overteach — it's Level 1 material.
0:17	8	Idea 3The agent loop	Perceive → Reason → Act → Observe → repeat (a.k.a. ReAct). Each pass gets closer.
0:25	12	⚡ EnergizerAgent or Not?	Sort tasks to AGENT / CHATBOT walls & defend. ✅ they can hear a task & classify it
0:37	35	🛠️ BuildAgent in Colab	Connect Gemini · give it a calculator + date tool · run the loop on "today's date & 1440×365". ✅ 🛠️ Act + 👀 Observe lines, correct answer
1:12	5	Wrap + exit	Exit: "3 parts of an agent?" Tease Session 2 (tools, prompting, memory).

Trim the engine recap to 3 min. Never cut the Colab build — it's the payoff.

If the agent answers with no tool call, re-run the config cell (system instruction dropped). Key errors = a stray space.

Facilitator Run-Sheet · Level 1.5 · Session 2 of 3

≈ 85 min Stations 4–6 · Deck: session-2 · Notebook: Session 2 · Energizer: Exact Instructions

Clock	Min	Block	Say / do — key cues
0:00	3	Welcome	Recap the loop. "Today we give the agent more power: tools, sharp prompts, memory."
0:03	9	Idea 1Tools & actions	A tool = a function it can call · a good tool has a clear name + description + schema. Examples: search, code, APIs.
0:12	8	Idea 2Prompting	The system prompt = job description · be specific, give examples, set the format. Vague → wrong tool.
0:20	7	Idea 3Memory	Short-term (context window) vs long-term (vector DB). Why: the window fills up & resets.
0:27	13	⚡ EnergizerExact Instructions	A "robot" follows written steps LITERALLY; teams rewrite to be unambiguous. ✅ they feel why specificity matters
0:40	35	🛠️ BuildTools + prompt + memory	Function-calling tools · vague-vs-sharp prompt · embed facts + retrieve (mini memory). ✅ right fact retrieved; sharp prompt wins
1:15	5	Wrap	Exit: "Why does an agent need memory?" Tease Session 3 (real-world agents + safety).

The memory (embeddings) part is the richest — protect it; trim the prompting build to a quick demo if needed.

Embedding model name errors — the notebook auto-picks one. Keep the "robot" energizer fast and funny.

Facilitator Run-Sheet · Level 1.5 · Session 3 of 3

≈ 88 min Stations 7–10 · Deck: session-3 · Notebook: Session 3 · Energizer: The Hidden Instruction

Clock	Min	Block	Say / do — key cues
0:00	3	Welcome	"Today: how real agents are structured, built, and kept safe."
0:03	10	Idea 1Architectures	ReAct · Chain of Thought · MCP ("USB-C for AI tools") · multi-agent teams.
0:13	8	Idea 2Building agents	Three paths: from scratch → function calling → frameworks (LangChain, CrewAI). Control vs speed.
0:21	13	⚡ EnergizerThe Hidden Instruction	"Agent" must obey only the coach — not sneaky instructions hidden in notes it reads. ✅ they catch a prompt injection
0:34	35	🛠️ BuildTraced agent + injection	A ReAct agent that logs every step · a prompt-injection demo (tricked → then defended). ✅ trace visible; agent resists injection
1:09	9	RecapTesting & safety	Agents are non-deterministic → log everything; sandbox tools (least privilege); protect privacy.
1:18	5	Finale 🏆	Celebrate finishing the agents track. Point to Level 2 to build a full project (Weeks 2–5).

Protect the prompt-injection demo — it's the memorable one. Trim architectures to ReAct + MCP if needed.

Keep injection payloads silly/harmless. Celebrate loudly — they've finished the whole agents track.