# Quiz — Session 7: Multi-agents & MCP **Program:** Applied AI — Intermediate Level · Instructor: Yann Isola **Format:** 10 multiple choice questions (MCQ). Only one correct answer per question. Recommended scale: 1 point per question. Correction commented at the end of the document. --- ### Question 1 What is the main argument for splitting a monolithic agent into several specialized agents? - A. Each agent necessarily requires a different AI (artificial intelligence) model - B. Separation of responsibilities: each agent has its own prompt system, its own tools and its own safeguards - C. Language models cannot handle more than three tools at a time - D. It’s always cheaper than a single agent ### Question 2 In the **orchestrator/worker** pattern, what is the role of the orchestrator? - A. He carries out all the tasks himself, the performers only serve as backup - B. It monitors the conformity of outputs and interrupts agents that drift - C. He breaks down the mission, delegates the sub-tasks to specialized agents, then assembles their results - D. It translates messages between agents who use different models ### Question 3 What is the key difference between the **pipeline** boss and the **orchestrator** boss? - A. The pipeline uses several models, the orchestrator only one - B. In the pipeline, the path is fixed in advance by the developer; with the orchestrator, it is an agent who dynamically decides on the decomposition - C. The pipeline is always faster than the orchestrator - D. The orchestrator only works with MCP (Model Context Protocol) tools ### Question 4 When is the pattern **debate/consensus** justified, despite its cost multiplied by the number of agents? - A. For all tasks, because it is always more reliable - B. Only for translation tasks - C. Never: it is a purely theoretical pattern - D. For high-stakes decisions, where the cost of an error far exceeds the cost of additional appeals ### Question 5 What distinguishes the **supervisor** from the **orchestrator**? - A. The orchestrator distributes the work; the supervisor controls quality and safety (he can interrupt or reassign) - B. The supervisor is a human, the orchestrator a model - C. The supervisor can only monitor one agent at a time - D. None: these are two names for the same boss ### Question 6 Blackboard architecture is: - A. A structured common space where each agent submits their contributions and is triggered when information that concerns them appears - B. A private message channel between exactly two agents - C. An error log reserved for the supervisor - D. The internal memory of the model, inaccessible to developers ### Question 7 What does MCP mean and what problem does it solve? - A. Multi-Chat Protocol: it allows multiple users to talk to the same model - B. Model Context Protocol: a universal standard that avoids coding N×M tailor-made integrations between models and tools — “the USB of AI” - C. Machine Control Program: it controls the physical servers of the data centers - D. Model Compression Protocol: it reduces the size of models to run them locally ### Question 8 In the MCP architecture, what are the roles of the Host, Client and Server? - HAS.Host = the datacenter; Customer = the human user; Server = the AI model - B. Host = the tool provider; Client = your application; Server = model - C. Host = your application (where the model lives); Client = the connector inside the host (one per server); Server = the program that exposes tools, resources and prompts - D. The three terms designate the same component seen from three angles ### Question 9 What is the difference between a **Tool** and a **Resource** MCP? - A. The Tool is free, the Resource requires payment - B. The Tool is an action invoked by the model; the Resource is a data read controlled by the application - C. The Tool works locally, the Resource only remotely - D. None: the two terms are interchangeable in the specification ### Question 10 Your agent runs on your workstation and must read local files via an MCP “filesystem” server installed on the same machine. Which transport is the natural choice? - A. HTTP+SSE (HyperText Transfer Protocol + Server-Sent Events), because it is the only transport defined by MCP - B. Bluetooth, provided for by the specification for local exchanges - C. stdio (standard input/output): the server is a local process launched by the host — simple, fast, everything stays on the machine - D. No transportation is necessary: the model reads directly from the hard drive --- --- ## Fixed commented **1 — B.** The separation of responsibilities is THE motivation: prompt system without contradictions, tools limited to need (least privilege, Session 5), targeted safeguards, debugging by agent. A is false: agents often share the same model. D is false: multi-agent generally costs *more* in calls — we choose it for robustness, not economy. **2 — C.** The orchestrator breaks down, delegates, assembles. He does not work in place of the performers (A), and quality/safety control is the role of the *supervisor* (B). **3 — B.** This is the structuring distinction of the session: pipeline = flow set by the developer (assembly line); orchestrator = decomposition decided at execution by an agent (project manager). Neither the speed (C) nor the number of models (A) define these patterns. **4 — D.** The debate multiplies the calls by the number of debaters: it is justified when the issue merits it (regulatory opinion, costly decision) — the independent responses overlap, and the divergences are themselves valuable information to be relayed to humans. **5 — A.** Orchestrator = distribution of work; supervisor = continuous control (quality, compliance, interruption). The two are often combined, particularly in regulated sectors. **6 — A.** The blackboard is a *structured* shared memory: sections, statuses, triggers. Course analogy: the whiteboard in the situation room — everyone writes what they know on it and reacts to what others have put there. **7 — B.** MCP = Model Context Protocol, proposed at the end of 2024 ⚠ by Anthropic and published as an open standard. Without standards: N models × M tools = N×M integrations. With MCP: N + M connectors. Hence the USB (Universal Serial Bus) analogy: a single socket for all devices. **8 — C.** The trio to know by heart. Beware of the trap of the word “server”: it is a *role* (capacity provider), often a simple small local program — not a datacenter machine.And the server does not contain a model: the intelligence remains on the host side. **9 — B.** All the difference lies in the trigger: Tool → the *model* decides to act (this is the tool calling of Session 5, standardized); Resource → the *application* decides which data to inject (RAG spirit, Session 4). Third primitive, for the record: Prompt → prompt template chosen by the *user*. **10 — C.** stdio = local (process launched by the host, dialog via standard input/output); HTTP+SSE = remote/shared. For a filesystem server on the same machine, stdio is the natural choice. D is dangerous in addition to being false: the model *never* accesses the disk directly — everything goes through controlled tools, this is precisely the principle of Sessions 5 to 7. --- ### Indicative scale - **9–10 /10:** excellent control — patterns, communication and MCP in place. - **7–8 /10:** solid; revisit the failed question, often the Tool/Resource or pipeline/orchestrator distinction. - **5–6 /10:** reread the Sequences B and E guide; repeat Exercise 2 (MCP architect). - **< 5/10:** resume the session with the interactive webpage (both tabs) before Session 8.