Unit 5: LLM-Driven Agents and Generative Workflows - Practice Quiz

1 What is the main role of an LLM in an agentic framework?

2 Which component typically carries out an action selected by an LLM agent?

3 What commonly guides an LLM agent toward its intended behavior?

4 In agent planning, what does chain-of-thought reasoning generally involve?

5 Why is step-based planning useful for an LLM agent?

6 What is the purpose of memory augmentation in an LLM agent?

7 Which type of information is most suitable for an agent's long-term memory?

8 What action does an agent perform when it searches stored memory for relevant facts?

9 Why might an LLM agent call a weather API?

10 What does API orchestration mean in an agentic workflow?

11 Which tool is most appropriate for an agent that must perform an exact arithmetic calculation?

12 What two activities are combined in the ReAct framework?

13 In a ReAct workflow, what is an observation?

14 What can a ReAct agent do after receiving an observation?

15 What is LangChain commonly used for?

16 What is a LangChain agent designed to do?

17 What is AutoGPT mainly known for demonstrating?

18 What does autonomous task delegation involve?

19 Why might a multi-agent system delegate a research subtask to a specialized agent?

20 Which example best represents an LLM agent used for decision support?

21 A customer-support agent must answer questions, query an order database, and request human approval before issuing large refunds. Which architecture best supports these requirements?

22 An LLM agent produces different action formats on repeated runs, causing downstream failures. What is the most appropriate improvement?

23 An agent receives the goal: "Prepare a competitor report and email it to the manager." Which planning approach is most effective?

24 A planning agent often continues executing after a required data source fails. Which mechanism would most improve its behavior?

25 A technical-support agent must retrieve earlier cases that are semantically similar even when users use different wording. Which memory design is most suitable?

26 A long-running assistant exceeds its context limit because it includes every previous message. What is the best solution?

27 An agent needs to remember a user's preferred report format across sessions but should not treat temporary task notes as permanent. What should the system do?

28 A weather agent calls an API with city names, but requests sometimes fail because required parameters are missing. Which change best addresses the problem?

29 An agent retries a timed-out payment API call, occasionally charging a customer twice. Which mechanism most directly prevents duplicate charges?

30 A scheduling agent can read calendars, create events, and delete events. How should tool permissions be configured for an agent that only proposes meeting times?

31 In a ReAct-style agent, what should normally happen after the agent selects and invokes a search tool?

32 A ReAct agent repeatedly calls the same unsuccessful tool with identical arguments. Which control is most appropriate?

33 A fact-checking agent must answer a question using a document search tool. Which behavior best reflects the ReAct framework?

34 In a LangChain agent, what is the primary role of a tool description?

35 Which scenario most clearly benefits from an AutoGPT-style autonomous loop rather than a single LLM call?

36 A manager agent delegates data collection, statistical analysis, and report writing to specialist agents. How should it reduce errors caused by task dependencies?

37 Two delegated agents produce conflicting recommendations based on different sources. What should the coordinator agent do?

38 A delegated sub-agent receives access to customer records even though it only formats a report. Which design principle has been violated?

39 An LLM agent ranks loan applications for staff review. Which design best supports responsible decision support?

40 A factory uses an LLM agent to recommend maintenance based on sensor summaries. What is the best response when required sensor data is missing?

41 An LLM agent can read customer records, issue refunds, and send email. Which architecture most effectively limits damage if the model is manipulated by an adversarial document?

42 An agent must choose between answering from its model, querying an expensive authoritative service, or abstaining. Which control strategy best accounts for both uncertainty and action cost?

43 A planning agent produces fluent multi-step rationales, but small changes in irrelevant wording cause different plans. Which intervention most directly improves plan reliability without treating the rationale as proof of correctness?

44 A partially observable task permits replanning after each action. Why can a fixed chain-of-thought plan generated before execution be inferior to a contingent policy?

45 An agent retrieves an old note stating that a deployment endpoint is active, although the endpoint was recently retired. Which memory design most directly addresses this failure?

46 A long-running agent compresses its episodic memory into summaries. Which invariant is most important for preserving its ability to audit consequential actions?

47 A user invokes a right-to-deletion request after their data has been incorporated into an agent's summaries and vector index. What is the most complete response?

48 An agent times out after submitting a payment request and cannot tell whether the server processed it. Which mechanism best prevents a retry from charging the customer twice?

49 An agent books a flight, hotel, and rental car through services that do not support a shared transaction. The hotel booking fails after the flight succeeds. Which orchestration pattern is most appropriate?

50 A web-search tool returns a page containing instructions telling the agent to reveal secrets and ignore its task. What is the strongest general mitigation?

51 In a ReAct loop, a tool observation contradicts a fact assumed in the preceding reasoning trace. What should the agent do?

52 A ReAct agent repeatedly alternates between two search queries without gaining information. Which termination rule best handles this while preserving useful exploration?

53 For a task with expensive tool calls and strict dependencies among actions, which modification to basic ReAct is most likely to reduce avoidable execution cost?

54 A LangChain-style agent keeps selecting a tool with semantically appropriate behavior but supplies malformed arguments. Which change most directly addresses the problem?

55 An AutoGPT-like system recursively creates goals and executes them unattended. Which addition most directly reduces the risk of endless self-generated work?

56 A coordinator delegates tasks represented by a directed acyclic graph. Task requires outputs from both and , while each of those requires . When may safely begin?

57 A manager agent can delegate database analysis to several worker agents with different trust levels. Which policy best limits privilege escalation through delegation?

58 Two worker agents produce incompatible answers from different data snapshots. What should a robust coordinator do before selecting either result?

59 A decision-support agent recommends whether to escalate cases to specialists. Errors have unequal costs, and the data distribution changes over time. Which evaluation design is most appropriate?

60 An automated decision-support system can either recommend an action or defer to a human. Let model error cost be , human-review cost be , and estimated error probability be . Ignoring other costs, when should it defer?