# Quiz — Session 6: The agentic loop **Program:** Applied AI — Intermediate Level · Instructor: Yann Isola **Format:** 10 multiple choice questions (MCQ). Only one correct answer per question, unless otherwise noted. Recommended scale: 1 point per question. Correction commented at the end of the document. --- ### Question 1 What is THE fundamental difference between a chatbot and an agent? - A. The agent uses a more powerful model than the chatbot - B. The chatbot stops after a response; the agent loops and decides itself when the task is accomplished - C. The agent works without a language model - D. The chatbot does not understand French ### Question 2 What is the correct order of agentic loop steps? - A. Act → perceive → observe → reflect - B. Think → observe → act → perceive - C. Perceive → reflect → act → observe - D. Observe → act → perceive → reflect ### Question 3 In the agentic loop, what does tool calling correspond to? - A. At the “perceive” stage - B. At the “think” stage - C. At the “act” stage — the tools are the agent’s hands - D. To a mechanism external to the loop ### Question 4 The TripDesk agent needs to book a flight. What does “planning” mean? - A. The flight schedule returned by the API (Application Programming Interface) - B. The decomposition, by the model, of the objective into sub-steps before acting — plan revisable along the way - C. The list of tools declared by the developer - D. The waiting time between two API calls ### Question 5 How does an agent “remember”, in iteration 4, the results obtained in iteration 1? - A. Thanks to a permanent database installed with the model - B. He doesn't remember: each iteration starts from scratch - C. Model weights are updated at each iteration - D. Everything is kept in the conversation history (the notepad / scratchpad), reread at each iteration ### Question 6 What happens to an agent's working memory (scratchpad) at the end of its execution? - A. It is erased: it is a memory limited to the current execution - B. It is automatically merged into the model weights - C. It remains available for all future missions, without limit - D. It is sent to the model editor for training ### Question 7 Which of the following are typical agent safeguards? **(multiple answers)** - A. A maximum number of iterations - B. A budget cap (API cost or time) - C. A human checkpoint (human-in-the-loop) before irreversible actions - D. Increase the temperature of the model to make it more creative ### Question 8 The `reserver_vol` tool returns the “steal complete” error. What is the behavior of an agent with good error recovery? - A. Immediately retry the same call, identically, until it goes through - B. Stop definitively and declare the mission impossible - C. Ignore the error and proceed to sending the confirmation - D. Read the error message, adapt your strategy (e.g. choose another flight) and try again differently ### Question 9 A user asks: “Translate this paragraph into Spanish. » What is the best architecture? - A. A complete agent with buckle, tools and guardrails -B.A simple prompt: one response is enough, everything else is over-engineering - C. A 6-step workflow with human validation at each step - D. Two agents who read each other ### Question 10 Why does the quality of the **error messages** returned by your tools directly influence the robustness of your agent? - A. Because detailed errors are required by regulations - B. Because the model adapts its strategy based on what it *reads*: an explicit message (“complete flight”) allows a bounce, an opaque code (“error 500”) does not - C. Because short messages cost more in tokens - D. This has no influence: only the power of the model matters --- --- ## Fixed commented **1 — B.** Agency is not a question of the power of the model (often it is *the same* model, answer A wrong): it is an architecture. The chatbot makes a turn then stops mechanically; the agent chains together actions and **decides itself** that the objective has been achieved. **2 — C.** Perceive (read the mission and the current state) → think (decide on the next action) → act (call a tool) → observe (read the result), then the loop starts again until the agent judges the task accomplished. **3 — C.** The tool call IS the “act” step. Without tools, the agent has no hands: its loop could only produce text. Callback Session 5: model *requests* the call; it's the application code that actually executes it. **4 — B.** The plan is produced by the model itself (often encouraged by the prompt system) and lives in the context. It is not fixed: like a GPS (Global Positioning System) which recalculates the route, the agent revises its plan when observations contradict it — p. ex. an agenda conflict discovered along the way. **5 — D.** Each tool call and each result is added to the conversation history, which the model rereads in full at each iteration. This is his working memory. Practical consequence: it is limited by the context window — an agent who is too talkative can saturate it. **6 — A.** Working memory is ephemeral: new mission = blank notepad. *Persistent* memory between sessions (preferences, profiles) is a separate mechanism, which must be designed explicitly. B and D are false: nothing is learned in the weights during execution. **7 — A, B and C.** The canonical trio: max iterations (against infinite loops), budget ceiling (against bleeding costs), human validation (against wrongly irreversible actions). D is a generation setting, not a guardrail — and more "creativity" is rarely what you want from an agent manipulating payments. **8 — D.** The complete reason: the tool fails → the model reads the error → adapts the strategy → retry **differently**. A is the classic anti-pattern (identical repetition — this is what the max iterations guardrail stops), B gives up too early, C is dangerous (confirming a reservation that does not exist!). **9 — B.** A translation is a one-turn task: no external data, no actions, no multi-step decisions. Golden rule: prompt < single tool < fixed workflow < agent — always the simplest weapon that accomplishes the mission. An agent here would add cost, latency and risk for zero benefit.**10 — B.** The “observe” step only has value if the observation is usable. The model can only adapt its strategy based on what is written in the context: a good error message actually contains the recovery path (“flight complete” → look for another flight; “quota exceeded, try again in 60 s” → wait). The robustness of an agent is conceived on the tools side as much as on the model side. --- **Indicative scale:** 9–10: excellent mastery · 7–8: solid, review missed points · 5–6: reread the guide and repeat exercise 1 · < 5: review the session with the interactive viewer before the next session.