AI Agents

Interactive Quiz - Only Have MCQs

Q1. (MCQ) An LLM is given access to a calculator tool and persistent memory. It can autonomously decide which tools to use, evaluate outcomes, and determine its own sequence of actions toward a goal. According to the course material, this system is best classified as:

A) A standard LLM with prompt engineering B) An Augmented LLM that has evolved into an agentic system C) A pure Augmented LLM with no agentic properties D) A multi-agent orchestration system

Answer: B

A) — Incorrect. A standard LLM with prompt engineering has no tools, no memory, and no autonomous action selection. The system described has all three.
B) — Correct. An Augmented LLM is not traditionally considered an agent because it lacks autonomous planning and goal-directed behavior. However, as it is granted persistent memory and the autonomy to dynamically select tools, evaluate outcomes, and determine its own sequence of actions in pursuit of a goal, it evolves into an agentic system, making the distinction increasingly blurred.
C) — Incorrect. A pure Augmented LLM enhances the base LLM's capabilities through external systems (tools and memory) but lacks autonomous goal-directed behavior. The system described goes beyond this by autonomously planning and selecting actions.
D) — Incorrect. A multi-agent system involves multiple coordinated agents. The description involves a single entity with tools and memory, not multiple agents.

Q2. (MSQ — Select ALL that apply) According to the classic agent definition from Russell & Norvig, an agent interacts with its environment through which components?

A) Sensors — used to observe the environment B) Actuators — tools used to interact with the environment C) Embeddings — used to encode the environment into vectors D) Effectors — the "brain" or rules deciding how to go from observations to actions

Answer: A, B, D

A) — Correct. Sensors are used to observe the environment — this is how the agent perceives its surroundings.
B) — Correct. Actuators are the tools used to interact with and act upon the environment.
C) — Incorrect. Embeddings are a machine learning technique for representing data as vectors. They're not part of the classic agent definition from Russell & Norvig — that framework predates modern embedding techniques.
D) — Correct. Effectors are the "brain" or rules deciding how to go from observations to actions — the decision-making component.

Q3. (MCQ) A team is building an AI agent prototype. They're debating whether to start with a small, cheap model to save costs or a large, capable model for quality. What is the recommended approach?

A) Start with the smallest model to minimize costs, then scale up only if quality is unacceptable B) Use a medium-sized model as a compromise between cost and quality for all tasks C) Build the prototype with the most capable model to establish a performance baseline, then swap in smaller models where acceptable D) Use different models for each task from the start, assigning complexity levels before testing

Answer: C

A) — Incorrect. Starting with the smallest model risks prematurely limiting the agent's abilities. You won't know if failures stem from the model's limitations or from poor agent design. You lose your quality baseline.
B) — Incorrect. A one-size-fits-all medium model doesn't establish a true performance ceiling and may be simultaneously too expensive for simple tasks and too weak for complex ones.
C) — Correct. The recommended approach is to build with the most capable model for every task to establish a performance baseline. From there, try swapping in smaller models to see if they still achieve acceptable results. This prevents premature limitation and lets you diagnose where smaller models succeed or fail.
D) — Incorrect. Pre-assigning complexity levels before any testing is guesswork. You need the baseline from the best model first to know which tasks actually require capability and which don't.

Q4. (MCQ) An LLM agent needs to answer a user's question, then update a CRM record, then send a confirmation email. Each intermediate step's output feeds back into the LLM for continued processing. This pattern of sequential LLM calls with autonomous tool selection describes:

A) A fixed agentic framework where tool order is predetermined B) A multi-agent decentralized handoff pattern C) An autonomous agent framework where the LLM chooses which tool to use and when D) Prompt chaining with no tool involvement

Answer: C

A) — Incorrect. In a fixed agentic framework, tools are used in a given order predefined by the developer. The description says the agent autonomously selects tools, which implies dynamic ordering.
B) — Incorrect. A decentralized handoff involves multiple agents transferring control to each other. This describes a single agent using multiple tools sequentially.
C) — Correct. LLM Agents are essentially sequences of LLM calls with autonomous selection of actions/tools. The output of intermediate steps is fed back into the LLM to continue processing. The agent decides the sequence: answer → update CRM → send email, selecting tools dynamically.
D) — Incorrect. Prompt chaining involves splitting tasks into subtask prompts, but the description explicitly involves tool usage (CRM updates, email sending), not just text transformations between prompts.

Q5. (MSQ — Select ALL that apply) Which of the following are the five fundamental components of AI Agents as defined in the course material?

A) Model — the LLM powering reasoning and decision-making B) Tools — external functions or APIs the agent can use C) Embeddings — vector representations for semantic search D) Memory — mechanisms to store, recall, and build upon past interactions

Answer: A, B, D

A) — Correct. The Model (LLM) is listed as component #1, powering the agent's reasoning and decision-making.
B) — Correct. Tools are listed as component #2 — external functions or APIs the agent can use to take action.
C) — Incorrect. Embeddings are not listed as a standalone agent component. They are a technique used within the Memory component (for vector database storage), but they are not one of the five fundamental components. The five are: Model, Tools, Instructions, Memory, and Planning/Reasoning.
D) — Correct. Memory is listed as component #4 — mechanisms that allow the agent to store, recall, and build upon past interactions and context.

Q6. (MCQ) An agent's conversation history has grown to 50,000 tokens, but the model's context window is only 8,192 tokens. The team cannot switch to a larger model. What is the recommended technique to maintain short-term memory?

A) Store the entire conversation in a vector database and retrieve relevant portions per turn B) Use another LLM to continuously summarize the conversation, keeping only the most vital information within the context window C) Truncate the conversation to the most recent 8,192 tokens, discarding everything older D) Split the conversation across multiple simultaneous model instances

Answer: B

A) — Incorrect. Storing in a vector database and retrieving relevant portions describes long-term memory via RAG, not short-term memory management. Short-term memory is about maintaining conversational continuity, not semantic retrieval.
B) — Correct. For models with smaller context windows or when conversation history is large, another LLM can be used to summarize the conversations that happened thus far. By continuously summarizing, you keep the conversation size small — reducing tokens while retaining only the most vital information.
C) — Incorrect. Hard truncation would lose all context from earlier in the conversation, potentially dropping critical information established at the start (like user identity, problem description, or decisions made). Summarization preserves the meaning even when the raw text is lost.
D) — Incorrect. Splitting across simultaneous instances would create fragmented contexts with no model having the complete picture. LLMs process a single context window per inference call — parallel instances don't share state.

Q7. (MCQ) The three types of tools an agent needs are Data, Action, and Orchestration. An agent that serves as a tool for another agent (e.g., a "Refund Agent" called by a "Manager Agent") falls under which type?

A) Data tool B) Action tool C) Orchestration tool D) It is not a tool — agents cannot be tools for other agents

Answer: C

A) — Incorrect. Data tools enable agents to retrieve context and information (query databases, read PDFs, search the web). An agent serving another agent is not a retrieval mechanism.
B) — Incorrect. Action tools enable agents to interact with systems to take actions (send emails, update records). While a sub-agent might trigger actions internally, the classification of an agent-as-tool specifically falls under Orchestration.
C) — Correct. Orchestration tools are explicitly defined as: "Agents themselves can serve as tools for other agents — see the Manager Pattern in the Orchestration section." A Refund Agent called by a Manager Agent is an orchestration tool.
D) — Incorrect. The material explicitly states that agents can serve as tools for other agents. This is a fundamental concept in multi-agent orchestration, particularly the Manager pattern.

Q8. (MCQ) An agent needs to use a calculator tool. To invoke it, the LLM generates a structured text output that matches the tool's API. What format is this output typically expected in?

A) Natural language sentences describing the desired calculation B) JSON-formatted strings that can be fed to a code interpreter C) Raw Python code that is directly executed by the operating system D) Binary-encoded function pointers

Answer: B

A) — Incorrect. Natural language is ambiguous and cannot be reliably parsed by a code interpreter. Tool invocation requires structured, deterministic formatting.
B) — Correct. To use a tool, the LLM generates text that fits the tool's API. The expected format is strings that can be formatted to JSON so they can easily be fed to a code interpreter. This structured approach enables reliable function calling.
C) — Incorrect. While some systems (like PAL) generate Python code, the general tool-use mechanism described here uses JSON-formatted function calls, not raw executable code. The code interpreter parses the JSON and dispatches the appropriate tool call.
D) — Incorrect. LLMs generate text tokens, not binary data. Binary function pointers are a systems programming concept unrelated to how LLMs invoke tools.

Q9. (MCQ) A developer has a single agent handling customer service. The agent's prompt contains 15 conditional if-then-else branches for different scenarios, and prompt templates are becoming unmanageable. The agent also frequently selects the wrong tool from 12 similar-sounding options. What should the developer do?

A) Add more tools to give the agent broader capabilities B) Switch to a smaller, faster model to reduce latency from the complex prompt C) Split the system into multiple agents, separating logical segments and overlapping tools into distinct specialized agents D) Remove all conditional branches and let the model figure out the logic autonomously

Answer: C

A) — Incorrect. Adding more tools would exacerbate the tool selection problem. The agent is already struggling with 12 similar tools — more would increase confusion.
B) — Incorrect. A smaller model would likely perform worse with complex conditional logic and ambiguous tool selection. The problem is architectural complexity, not latency.
C) — Correct. Two practical guidelines for splitting agents are identified: (1) Complex logic — when prompts contain many conditional statements and templates get difficult to scale, divide each logical segment across separate agents; (2) Tool overload — when tools are similar or overlapping and improving clarity doesn't help, use multiple agents. Both conditions are met here.
D) — Incorrect. Removing conditional branches without replacing them with structured guidance would make the agent even more unpredictable. The model needs clear instructions — the solution is distributing that complexity across specialized agents, not removing it.

Q10. (MSQ — Select ALL that apply) Which of the following are valid exit conditions for an agent's execution loop (the "run")?

A) A final-output tool is invoked, defined by a specific output type B) The model returns a response without any tool calls C) An error occurs during execution D) The agent has been running for more than 5 seconds

Answer: A, B, C

A) — Correct. A final-output tool being invoked (defined by a specific output type) is explicitly listed as an exit condition for the agent loop.
B) — Correct. The model returning a response without any tool calls (e.g., a direct user message) is another explicit exit condition — the agent has nothing more to do.
C) — Correct. Errors are listed among common exit conditions alongside tool calls, structured output, and maximum turns.
D) — Incorrect. Time-based limits (wall-clock duration) are not listed as an exit condition. The listed conditions are: final-output tool invocation, no tool calls in the response, errors, and reaching a maximum number of turns (not seconds).

Q11. (MCQ) The Cognitive Architectures for Language Agents paper couples four types of memory to LLM Agents. The material explains that semantic memory (facts about the world) might be stored differently from working memory (current and recent circumstances). Why is this differentiation useful for building agentic frameworks?

A) It allows using different storage backends optimized for each memory type's access patterns and content B) It eliminates the need for a context window entirely C) It ensures the agent never hallucinates by separating facts from current context D) It reduces the total number of tokens the agent needs to process to zero

Answer: A

A) — Correct. The differentiation helps in building agentic frameworks because different types of information have different storage, retrieval, and persistence requirements. Semantic memory (stable world facts) might be stored in a different database than working memory (transient current circumstances), allowing each to be optimized for its specific access patterns.
B) — Incorrect. The context window is still essential for working memory and the LLM's immediate processing. Memory type differentiation complements the context window, not replaces it.
C) — Incorrect. Separating memory types improves organization and retrieval efficiency, but it doesn't eliminate hallucination. The LLM can still hallucinate regardless of how its external memory is organized.
D) — Incorrect. The agent still processes tokens — memory type differentiation affects what gets loaded into the context and from where, but the LLM still processes tokens during inference.

Q12. (MCQ) In the Manager pattern of multi-agent orchestration, how does the central "manager" agent interact with specialized agents?

A) By handing off full execution control to a specialized agent, which then interacts directly with the user B) By delegating tasks to specialized agents via tool calls, then synthesizing their results into a cohesive interaction C) By running all specialized agents in parallel and selecting the best output D) By splitting the user's message into tokens and distributing them across agents

Answer: B

A) — Incorrect. This describes the decentralized pattern, where agents hand off execution to peers who then interact directly with the user. In the Manager pattern, the manager retains control and user access.
B) — Correct. The Manager pattern empowers a central LLM to orchestrate a network of specialized agents seamlessly through tool calls. Instead of losing context or control, the manager intelligently delegates tasks, then synthesizes results into a cohesive interaction. This ensures a smooth, unified user experience with specialized capabilities available on-demand.
C) — Incorrect. Parallel execution with output selection isn't the Manager pattern — the manager sequentially delegates specific tasks to the right agent at the right time, not all at once.
D) — Incorrect. Token-level distribution across agents is not a real orchestration pattern. Agents operate on semantic tasks, not individual tokens.

Q13. (MCQ) A company is considering building an AI agent to automate invoice processing. Their invoices follow a rigid template with fixed fields, consistent formatting, and deterministic validation rules with no exceptions. Should they build an AI agent?

A) Yes — all automation benefits from AI agents B) Yes — the agent could learn to handle the template over time C) No — the use case is deterministic with no complex judgment, unstructured data, or difficult-to-maintain rules, so a simpler solution is likely sufficient D) No — AI agents cannot process structured documents

Answer: C

A) — Incorrect. Not all automation requires an AI agent. The material explicitly warns: if your use case doesn't fall into the categories of complex decision-making, difficult-to-maintain rules, or heavy reliance on unstructured data, a simpler deterministic solution may be sufficient.
B) — Incorrect. There's nothing to "learn" — the template is rigid and fixed. An agent's strengths (nuanced judgment, handling exceptions, interpreting natural language) are wasted on deterministic, template-based processing.
C) — Correct. The three categories warranting an AI agent are: (1) complex decision-making with nuanced judgment, (2) difficult-to-maintain rules, and (3) heavy reliance on unstructured data. Rigid templates with fixed fields and deterministic rules match none of these — a traditional rules-based or template-parsing solution would be simpler, cheaper, and more reliable.
D) — Incorrect. AI agents can process structured documents — the question is whether they should. For rigid, deterministic processing, they add unnecessary complexity.

Q14. (MCQ) In the decentralized orchestration pattern, a customer service agent determines that the user needs a refund and transfers control to a specialized Refund Agent. The Refund Agent now interacts directly with the user. This transfer is called a:

A) Tool call B) Handoff C) Manager delegation D) Context window swap

Answer: B

A) — Incorrect. While handoffs are technically implemented as a type of tool/function in the Agents SDK, the specific term for this one-way transfer of execution control between peer agents is a "handoff," not a generic tool call. Tool calls in the Manager pattern serve a different architectural purpose (delegation with result synthesis, not execution transfer).
B) — Correct. In the decentralized pattern, agents can "handoff" workflow execution to one another. A handoff is a one-way transfer where an agent delegates to another agent. Execution immediately starts on the new agent, and the latest conversation state is transferred. The original agent does not need to remain involved.
C) — Incorrect. Manager delegation involves a central manager orchestrating specialized agents via tool calls while retaining control and user access. In this scenario, the original agent gives up control — that's a decentralized handoff, not manager delegation.
D) — Incorrect. "Context window swap" is not a defined concept. The conversation state is transferred alongside the handoff, but this is part of the handoff mechanism, not a separate operation.

Q15. (MSQ — Select ALL that apply) Which of the following are best practices for writing agent instructions?

A) Use existing operating procedures, support scripts, or policy documents to create LLM-friendly routines B) Keep instructions high-level and abstract to give the agent maximum creative freedom C) Define clear actions so every step corresponds to a specific action or output D) Capture edge cases with conditional steps for when users provide incomplete information or unexpected questions

Answer: A, C, D

A) — Correct. When creating routines, use existing operating procedures, support scripts, or policy documents. In customer service, routines can roughly map to individual knowledge base articles.
B) — Incorrect. The exact opposite is recommended. Instructions should prompt agents to break down tasks into smaller, clearer steps from dense resources to minimize ambiguity. Abstract, high-level instructions leave room for errors in interpretation.
C) — Correct. Every step in your routine should correspond to a specific action or output. Being explicit about the action (and even the wording of user-facing messages) leaves less room for interpretation errors.
D) — Correct. Real-world interactions create decision points. A robust routine anticipates common variations and includes instructions on how to handle them with conditional steps or branches.

Q16. (MCQ) Planning in LLM Agents involves breaking a task into actionable steps. During execution, the agent finds that Step 3 of its plan failed. What does the planning capability allow the agent to do?

A) Restart the entire task from Step 1 with the same plan B) Skip Step 3 and continue to Step 4 without adjustment C) Iteratively reflect on past behavior and update the current plan D) Terminate execution and report failure to the user

Answer: C

A) — Incorrect. Restarting from scratch wastes all progress from Steps 1-2 and doesn't address why Step 3 failed. Planning enables adaptation, not blind repetition.
B) — Incorrect. Skipping a failed step without adjusting the plan could propagate errors downstream. If Step 4 depends on Step 3's output, blindly continuing would produce incorrect results.
C) — Correct. The plan allows the model to iteratively reflect on past behavior and update the current plan if necessary. When Step 3 fails, the agent can reassess what went wrong, adjust the remaining steps, choose alternative approaches, or modify its strategy — this is the core value of planning.
D) — Incorrect. Immediate termination on first failure would make the agent fragile and useless for complex tasks. The planning capability specifically enables resilience through reflection and adaptation.

Q17. (MCQ) Reasoning behavior in LLMs can be enabled through two broad approaches. Which are they?

A) Increasing the model's context window and adding more tools B) Fine-tuning the LLM on reasoning data, or using specific prompt engineering techniques like Chain-of-Thought C) Storing reasoning traces in a vector database and retrieving them via RAG D) Using a manager agent to provide reasoning instructions to a worker agent

Answer: B

A) — Incorrect. A larger context window provides more space for information but doesn't inherently enable reasoning behavior. Tools extend capabilities but don't create the reasoning process itself.
B) — Correct. Reasoning behavior can be enabled by roughly two choices: fine-tuning the LLM (e.g., giving it datasets with thought-like examples, or having it discover its own thinking process like DeepSeek-R1 with reward-guided training) or specific prompt engineering (e.g., Chain-of-Thought with examples, or zero-shot CoT with "Let's think step-by-step").
C) — Incorrect. RAG retrieves external knowledge, not reasoning processes. You might retrieve facts via RAG, but the reasoning behavior itself must come from the model's training or prompting.
D) — Incorrect. Multi-agent delegation is an orchestration pattern, not a mechanism for enabling reasoning. The worker agent still needs its own reasoning capability to process the delegated task.

Q18. (MCQ) A single-agent system is handling 6 distinct customer service workflows using prompt templates with policy variables. A new workflow is added. Rather than creating an entirely new agent or rewriting the prompt, the developer updates a variable in the template. This approach is an example of:

A) Multi-agent orchestration via the manager pattern B) Using prompt templates with policy variables to manage complexity within a single agent C) Decentralized handoff between specialized agents D) Long-term memory retrieval from a vector database

Answer: B

A) — Incorrect. The manager pattern involves multiple agents coordinated by a central manager. This scenario describes a single agent adapting its behavior via template variables.
B) — Correct. An effective strategy for managing complexity without switching to a multi-agent framework is to use prompt templates. Rather than maintaining numerous individual prompts, a single flexible base prompt accepts policy variables. As new use cases arise, you can update variables rather than rewriting entire workflows — significantly simplifying maintenance and evaluation.
C) — Incorrect. No handoff or second agent is involved. The single agent adapts its behavior through template variable changes.
D) — Incorrect. Template variables are injected at prompt construction time, not retrieved from a vector database. This is a prompt engineering technique, not a memory retrieval mechanism.

Q19. (MCQ) Long-term memory in an LLM agent is implemented by storing all previous interactions in an external vector database. When a new prompt arrives, the system embeds it and finds the most relevant past interactions. This technique is referred to as:

A) Context window expansion B) Conversation summarization C) Retrieval-Augmented Generation (RAG) D) Short-term memory buffering

Answer: C

A) — Incorrect. Context window expansion would mean increasing the model's token limit. Storing interactions externally doesn't change the model's context window — it supplements it.
B) — Incorrect. Conversation summarization is a short-term memory technique that compresses recent dialogue to fit within the context window. It doesn't involve a vector database or embedding-based retrieval.
C) — Correct. Storing previous interactions, embedding them, and retrieving the most relevant ones based on similarity to a new prompt is explicitly identified as Retrieval-Augmented Generation (RAG). This is the common technique for enabling long-term memory in LLM Agents.
D) — Incorrect. Short-term memory buffering refers to using the context window itself to hold recent conversation history. The vector database approach described here is long-term memory — it spans across sessions and stores information beyond the context window's capacity.

Q20. (MCQ) A framework uses declarative graphs where developers must explicitly define every branch, loop, and conditional upfront using nodes (agents) and edges (handoffs). The Agents SDK takes a different approach. What is it?

A) A visual drag-and-drop interface for building agent workflows B) A code-first approach where developers express workflow logic using familiar programming constructs without pre-defining the entire graph C) A no-code platform that automatically infers the optimal graph structure D) A static configuration file that defines all possible agent interactions in YAML

Answer: B

A) — Incorrect. No visual drag-and-drop interface is described. The contrast is between declarative graph definition and code-first flexibility, not between code and visual tools.
B) — Correct. The Agents SDK adopts a more flexible, code-first approach. Developers can directly express workflow logic using familiar programming constructs without needing to pre-define the entire graph upfront, enabling more dynamic and adaptable agent orchestration. This contrasts with declarative frameworks that require learning specialized domain-specific languages.
C) — Incorrect. The SDK doesn't automatically infer graph structures — it gives developers programmatic control to define workflows imperatively in code, adapting dynamically as needed.
D) — Incorrect. Static YAML configuration would be even more rigid than declarative graphs. The SDK's advantage is dynamic flexibility through code, not static configuration.

Q21. (MCQ) An LLM agent adapts to the Augmented LLM framework for interacting with its environment. Through what modality does it primarily observe its environment, and through what mechanism does it act?

A) It observes through image inputs and acts through physical actuators B) It observes through textual input and acts through its use of tools C) It observes through sensor data streams and acts through API calls exclusively D) It observes through embeddings retrieved from a vector database and acts through natural language responses only

Answer: B

A) — Incorrect. While multimodal models exist, the material specifies that LLMs are generally textual models. Physical actuators belong to robotic agents, not the LLM agent framework described here.
B) — Correct. Using the Augmented LLM, the Agent can observe the environment through textual input (as LLMs are generally textual models) and perform certain actions through its use of tools (like searching the web). This maps the classic agent framework (sensors → effectors → actuators) to the LLM paradigm (text input → reasoning → tool use).
C) — Incorrect. "Sensor data streams" describes physical/IoT agents. LLM agents receive textual input, and their actions go beyond just API calls — they include any tool use the system is configured for.
D) — Incorrect. Embeddings from vector databases are one source of information (long-term memory), not the primary observation modality. And agents act through tools, not just natural language responses — tool-mediated actions (database updates, emails, calculations) are core to agentic behavior.