Agentic Systems quiz

Question 1

What is an LLM agent and what distinguishes it from a standard chatbot?

Accepted Answer

Agents autonomously plan, use tools, observe results, and iterate; chatbots respond to single queries

Answer

Agents are identical to chatbots but with a larger context window

Answer

Agents only work with specialized fine-tuned models, not standard LLMs

Answer

Agents cannot use external tools; they rely entirely on parametric knowledge

Question 2

What is function calling (tool use) in LLMs and how is it implemented?

Accepted Answer

Function calling has the LLM output structured JSON tool specs; the host executes the function and returns results

Answer

Function calling requires the LLM to execute the function itself without host application involvement

Answer

Function calling is only possible with models fine-tuned on code generation tasks

Answer

Function calling requires FP8 quantization for reliable JSON parsing

Question 3

What are the three main types of memory in LLM agent architectures?

Accepted Answer

Agent memory types: in-context (working), external (vector DB retrieval), and parametric (model weights)

Answer

LLM agents only support in-context memory; external storage is not possible

Answer

Parametric memory can be updated at inference time without retraining

Answer

In-context memory is unlimited in modern long-context models

Question 4

What is multi-agent orchestration and what are the main communication patterns?

Accepted Answer

Multi-agent patterns include sequential, hierarchical, debate, and parallel fan-out; implemented in LangGraph/AutoGen

Answer

Multi-agent systems can only use sequential pipelines for communication

Answer

Multi-agent systems require all agents to use the same underlying LLM

Answer

Orchestrator agents cannot delegate to other orchestrator agents

Question 5

What is planning in agentic systems and what are the main planning approaches?

Accepted Answer

Planning approaches include Plan-and-Execute, ReAct, Tree-of-Thought, and MCTS — depending on task complexity

Answer

LLM agents can only use one planning approach per task

Answer

MCTS planning requires the environment to have deterministic transitions

Answer

Planning is only necessary for tasks with more than 100 steps

Question 6

What is the agent control plane vs data plane separation?

Accepted Answer

Control plane handles orchestration and routing; data plane handles LLM inference and tool execution

Answer

Control plane handles LLM inference; data plane handles tool routing

Answer

Control and data planes must run on the same server for agent consistency

Answer

Data plane is responsible for deciding which agent runs next

Question 7

What is the HNSW (Hierarchical Navigable Small World) index and why is it preferred for ANN search?

Accepted Answer

HNSW uses a multi-layer graph for O(log N) approximate nearest neighbor search with high recall

Answer

HNSW achieves O(1) search by using a hash table for vector lookup

Answer

HNSW requires retraining the index when new vectors are added

Answer

HNSW is slower than IVF for all query types

Question 8

What is the difference between zero-shot, few-shot, and fine-tuned tool use?

Accepted Answer

Zero-shot uses descriptions only; few-shot adds examples; fine-tuned models are explicitly trained on tool use

Answer

Fine-tuned tool use is always worse than few-shot because it reduces flexibility

Answer

Zero-shot tool use requires the model to have been trained on the specific tool's API

Answer

Fine-tuned tool use eliminates the need for tool descriptions in the system prompt

Question 9

What is computer use in AI agents and what safety challenges does it introduce?

Accepted Answer

Computer use introduces prompt injection via screen, irreversible actions, and credential exposure risks

Answer

Computer use agents are completely safe because they cannot access the network

Answer

Computer use requires specialized hardware not available on consumer machines

Answer

Computer use agents only work on web browsers, not desktop applications

Question 10

What is a critic agent and how is it used in multi-agent debate?

Accepted Answer

Critic agents review and challenge proposer outputs; multi-agent debate converges to more accurate answers

Answer

Critic agents are identical to reward models in RLHF training

Answer

Critic agents always produce worse results than single-agent systems

Answer

Multi-agent debate is only applicable to math problems

Question 11

What is SWE-Bench and what does it evaluate?

Accepted Answer

SWE-Bench evaluates agents on resolving real GitHub issues by generating code patches that pass tests

Answer

SWE-Bench evaluates agents on generating documentation for Python repositories

Answer

SWE-Bench only evaluates code generation, not bug detection or test writing

Answer

SWE-Bench tasks require agents to train new models, not write code

Question 12

What is the sandbox environment and why is it critical for safe agent code execution?

Accepted Answer

Sandboxes isolate agent code execution to prevent network access, file system damage, and resource abuse

Answer

Sandboxes are not needed when using well-aligned LLMs like Claude

Answer

Sandboxes prevent agents from executing any code, limiting functionality

Answer

Sandboxes are only needed for agents writing C/C++ code, not Python

Question 13

What is GAIA (General AI Assistants) benchmark and what does it measure?

Accepted Answer

GAIA benchmarks real-world multi-step tasks requiring web search, tool use, and document analysis

Answer

GAIA evaluates LLM knowledge without allowing any tool use

Answer

GAIA is only used to evaluate code generation agents

Answer

GAIA Level 3 tasks are solved by current agents with >90% accuracy

Question 14

What is the task decomposition problem in planning agents?

Accepted Answer

Decomposition challenges include subtask ordering, under/over-decomposition, dependencies, and failure re-planning

Answer

Task decomposition always produces independent subtasks that can run in parallel

Answer

Task decomposition is only needed for tasks with more than 100 steps

Answer

LLM agents decompose tasks by only splitting the prompt into multiple parts

Question 15

What is the concept of agent trajectory and why is it important for evaluation?

Accepted Answer

Agent trajectory is the full sequence of observations, thoughts, and actions — essential for evaluation and training

Answer

Trajectory is only relevant for evaluation, not for training new agents

Answer

Trajectories are only stored for debugging and cannot be used as training data

Answer

A shorter trajectory always indicates a better-performing agent

Question 16

What is the difference between a prompt template and a prompt chain in agentic workflows?

Accepted Answer

Prompt templates define single reusable prompts; chains sequence multiple LLM calls with outputs as inputs

Answer

Prompt chains always require multiple LLM models for different steps

Answer

Prompt chains are deprecated in favor of single mega-prompts in modern frameworks

Answer

Prompt chains cannot branch conditionally based on intermediate LLM outputs

Question 17

What is self-consistency sampling in agent reasoning and when is it applied?

Accepted Answer

Self-consistency generates K reasoning paths and uses majority vote for improved accuracy

Answer

Self-consistency uses a single generation path with deterministic decoding

Answer

Self-consistency requires a trained verifier to score reasoning paths

Answer

Self-consistency only works when temperature is set to 0

Question 18

What is the lost-in-the-middle problem and how does it affect agent context management?

Accepted Answer

Lost-in-the-middle: LLMs have poor recall for content in the middle of long contexts; beginning/end are better recalled

Answer

Lost-in-the-middle means the model loses track of the task when contexts are too short

Answer

Lost-in-the-middle is solved by using Flash Attention for long sequences

Answer

Lost-in-the-middle only affects models with context windows under 4096 tokens

Question 19

What is agentic loop termination and what are common failure modes?

Accepted Answer

Agent failure modes include infinite loops, goal drift, premature termination, and tool failure cascades

Answer

Agents always terminate naturally without requiring explicit stopping conditions

Answer

Setting max_steps prevents all agent failure modes in production systems

Answer

Agent loops only fail when the LLM makes a factual error in its reasoning

Question 20

What is the subagent pattern in hierarchical multi-agent systems?

Accepted Answer

Orchestrator decomposes tasks and spawns specialized subagents; enables parallel execution and specialization

Answer

Subagents must use the same LLM as the orchestrator for consistency

Answer

Subagents cannot use tools; only the orchestrator can call external functions

Answer

Hierarchical agent systems are slower than single-agent systems for all task types

Question 21

What is the observability stack for production agentic systems?

Accepted Answer

Production observability needs trace logging, span correlation, cost tracking, error monitoring, and eval pipelines

Answer

Agent observability only requires logging the final output of each agent run

Answer

LangSmith and Langfuse are identical tools with different pricing

Answer

Observability is not needed for agents using Claude, which has built-in logging

Question 22

What is the principle of minimal footprint for agentic AI safety?

Accepted Answer

Minimal footprint: agents request minimum necessary permissions and prefer reversible actions to limit risk

Answer

Minimal footprint means agents should request maximum permissions upfront to avoid interruptions

Answer

Minimal footprint principle only applies to agents with internet access

Answer

Minimal footprint is a performance optimization to reduce agent API calls

Question 23

What is tool result injection and how does it update agent context?

Accepted Answer

Tool results are injected as user-role messages with tool_result type, correlated by tool_use_id

Answer

Tool results are injected as system prompt updates, not user messages

Answer

Tool results are automatically appended to the assistant message that called the tool

Answer

Tool results must always be formatted as JSON regardless of tool type

Question 24

What is the role of system prompts in defining agent personas and capabilities?

Accepted Answer

System prompts define agent role, scope, tools, and constraints with operator-level authority above user messages

Answer

System prompts are visible to end users and can be modified by them

Answer

System prompts are only used for safety filtering, not capability definition

Answer

System prompts are processed after user messages and can be overridden by them

Question 25

What is retrieval reranking and why is it used in RAG pipelines?

Accepted Answer

Reranking uses a cross-encoder on top-K bi-encoder results for more accurate relevance scoring

Answer

Reranking replaces the bi-encoder in RAG pipelines for all retrieval steps

Answer

Reranking is only needed for datasets with more than 1 million documents

Answer

Cross-encoder reranking is faster than bi-encoder retrieval for all corpus sizes

Question 26

What are guardrails in agentic AI systems and how are they implemented?

Accepted Answer

Guardrails intercept agent I/O to enforce safety policies; implemented via LLM classifiers or rule-based filters

Answer

Guardrails replace RLHF alignment and are the only safety mechanism needed

Answer

Guardrails are only applicable to agents with internet access

Answer

Input guardrails are applied after the agent generates a response

Test yourself