Agentic AI and tool use quiz

Question 1

In the ReAct (Reasoning + Acting) agent paradigm, what distinguishes it from a pure chain-of-thought approach?

Accepted Answer

It interleaves reasoning with tool actions and observations that are fed back into context

Answer

It uses a larger model than chain-of-thought

Answer

It disables the model's internal reasoning entirely

Answer

It only works for math problems

Question 2

When defining a tool for an LLM via function calling, why is a JSON Schema (with types, required fields, and descriptions) preferred over a free-text description alone?

Accepted Answer

It enables structured argument generation plus runtime validation before execution

Answer

It makes the model respond faster

Answer

It removes the need for any system prompt

Answer

It guarantees the tool never errors

Question 3

What problem does the Model Context Protocol (MCP) primarily solve for agentic systems?

Accepted Answer

A standardized client-server protocol for connecting LLMs to tools and data sources

Answer

A new model architecture that replaces transformers

Answer

A vector database optimized for embeddings

Answer

A fine-tuning method for tool selection

Question 4

In the Reflexion framework, how does an agent improve across attempts at a task?

Accepted Answer

It writes verbal self-reflections to memory and reuses them on later attempts

Answer

It backpropagates gradients after each episode

Answer

It increases the temperature on retries

Answer

It swaps to a larger model after failure

Question 5

An agent must call three independent read-only APIs whose results don't depend on each other. What is the recommended tool-calling strategy?

Accepted Answer

Emit all three as parallel tool calls in one turn

Answer

Call them strictly one at a time across three turns

Answer

Merge them into a single prompt with no tools

Answer

Fine-tune the model to memorize the API outputs

Question 6

What is a robust termination condition for an agent's reasoning-acting loop to prevent runaway execution?

Accepted Answer

A combination of max-step budget, cost ceiling, final-answer detection, and loop detection

Answer

Only the model's self-reported 'done' signal

Answer

A fixed 60-second timer with no step limit

Answer

Waiting until the context window is completely full

Question 7

Why is prompt injection a heightened risk specifically in agents that fetch and act on external content?

Accepted Answer

Untrusted fetched content can inject instructions that hijack the agent's real-world tool actions

Answer

External content increases token costs only

Answer

It only affects the model's grammar, not its actions

Answer

Injection is impossible once a system prompt is set

Question 8

In a multi-agent system, what is the key advantage of an orchestrator-worker (supervisor) pattern over a flat peer-to-peer agent mesh?

Accepted Answer

Centralized routing and aggregation improve control, debugging, and budget enforcement

Answer

It always uses fewer tokens regardless of task

Answer

It eliminates the need for any individual agent prompts

Answer

Workers can never communicate, improving privacy

Question 9

What does 'agentic RAG' add compared to classic single-shot retrieval-augmented generation?

Accepted Answer

Iterative, agent-controlled retrieval with query reformulation and verification

Answer

A larger embedding model only

Answer

Replacing the vector store with a SQL database

Answer

Disabling retrieval to rely on parametric memory

Question 10

How should an agent runtime handle a tool that returns an error (e.g., HTTP 500 or invalid args)?

Accepted Answer

Feed the error back as a tool observation and apply bounded retries with backoff

Answer

Immediately terminate the agent on any error

Answer

Hide the error and return an empty result silently

Answer

Automatically escalate temperature to 1.5

Question 11

In Anthropic's tool-use API, what role does the `tool_result` content block play in a conversation?

Accepted Answer

It returns a tool's output to the model, linked by tool_use_id, to continue the turn

Answer

It defines the JSON schema of available tools

Answer

It stores long-term memory across sessions

Answer

It sets the model's system prompt

Question 12

What is the primary purpose of prompt caching in long-running agentic workflows?

Accepted Answer

Reuse a stable prefix's processed state to cut input cost and latency on repeated calls

Answer

Permanently store conversations for training

Answer

Cache the model's weights on the GPU

Answer

Compress output tokens to save bandwidth

Question 13

Why might excessive numbers of available tools (e.g., 100+) degrade an agent's reliability, and what mitigates it?

Accepted Answer

It causes selection errors and context bloat; fix with tool retrieval/routing to a relevant subset

Answer

More tools always improve accuracy, no mitigation needed

Answer

Tools must be alphabetized to work

Answer

The only fix is using a smaller model

Question 14

What distinguishes 'plan-and-execute' agents from purely reactive (ReAct-style) agents?

Accepted Answer

They produce an explicit upfront plan and execute/re-plan, vs deciding one action at a time

Answer

They never use tools

Answer

They require reinforcement learning to plan

Answer

They cannot handle multi-step tasks

Question 15

In agent memory design, what is the difference between semantic and episodic memory?

Accepted Answer

Semantic stores generalized facts; episodic stores specific past experiences/events

Answer

Semantic is short-term; episodic is the system prompt

Answer

They are identical with different names

Answer

Episodic memory holds the model weights

Question 16

What is the main reliability concern with letting an agent generate and execute arbitrary code as a tool, and a common mitigation?

Accepted Answer

Unsafe side effects; mitigate with a resource-limited, isolated sandbox

Answer

Code is always safe so no mitigation is needed

Answer

The only risk is slow output formatting

Answer

Mitigate by raising the token limit

Question 17

When evaluating an agent, why are trajectory-level metrics important in addition to final-answer accuracy?

Accepted Answer

They expose efficiency, cost, and safety of the path, not just whether the answer was right

Answer

They are only useful for training data

Answer

Final accuracy already captures everything

Answer

They measure GPU temperature

Question 18

What is a 'hallucinated tool call' and how can a runtime defend against it?

Accepted Answer

Invoking a non-existent tool or invalid args; defend with strict validation and error feedback

Answer

A tool that runs too slowly; defend by caching

Answer

The model refusing to call tools; defend with retries

Answer

A duplicate tool name; defend by renaming the model

Question 19

In LangGraph, why is modeling an agent as a state graph advantageous over a simple while-loop?

Accepted Answer

Explicit nodes/edges enable checkpointing, interrupts, resumption, and deterministic branching

Answer

It runs the model on a GPU automatically

Answer

It removes the need for tools

Answer

Graphs are always faster than loops

Question 20

What is the purpose of a 'human-in-the-loop' approval gate in an agentic workflow, and where should it be placed?

Accepted Answer

Require approval before high-risk/irreversible tool actions to bound error impact

Answer

Approve every single token the model generates

Answer

Only review the final answer, never tool calls

Answer

Replace all tools with human labor

Question 21

Why should tool outputs be treated as untrusted 'data' rather than 'instructions' inside an agent's context?

Accepted Answer

Because they may contain injected instructions that hijack the agent if obeyed

Answer

Because data is always longer than instructions

Answer

Because instructions cost more tokens

Answer

Because the model cannot read tool outputs otherwise

Question 22

What does the τ-bench (tau-bench) benchmark evaluate in agentic systems?

Accepted Answer

Multi-turn tool-using task completion with policy adherence and reliability (pass^k)

Answer

Image classification accuracy

Answer

Raw token generation speed

Answer

Embedding retrieval recall only

Question 23

For long-horizon agents that exceed the context window, what is a common context-management strategy?

Accepted Answer

Summarize/compact old context and offload detail to retrievable external memory

Answer

Always switch to a model with infinite context

Answer

Drop the system prompt first

Answer

Repeat the full history every turn unchanged

Question 24

What temperature setting is generally recommended for the tool-selection/decision steps of a production agent, and why?

Accepted Answer

Low/near-zero temperature for deterministic, schema-correct tool decisions

Answer

High temperature to explore more tools

Answer

Always exactly 1.0 for all steps

Answer

Temperature has no effect on tool calls

Question 25

In computer-use / GUI agents, why is a perception-action loop with screenshots more challenging than text-based tool calling?

Accepted Answer

Visual grounding to pixels and coordinate actions is error-prone, slow, and partially observable

Answer

Screenshots are smaller than text so it is easier

Answer

GUIs always expose a clean JSON schema

Answer

It only differs in the model size required

Question 26

What is the role of an 'output verifier' or critic step in improving agent answer quality?

Accepted Answer

Independently check the output against task/evidence and trigger revision on failure

Answer

Translate the output into another language

Answer

Increase generation temperature

Answer

Cache the output for reuse

Question 27

Why is observability/tracing (e.g., spans for each LLM call and tool call) especially critical for agents versus single-shot LLM apps?

Accepted Answer

Multi-step nondeterministic behavior makes per-step traces essential to diagnose failures and cost

Answer

Agents never fail so tracing is optional

Answer

Tracing replaces the need for evaluation

Answer

It only matters for the UI layer

Question 28

What is the key idea behind giving an agent a 'scratchpad' or dedicated thinking space separate from its final answer?

Accepted Answer

Externalizing intermediate reasoning improves multi-step quality and keeps the final answer clean

Answer

It stores user passwords securely

Answer

It is only a UI animation

Answer

It disables tool calling

Question 29

In a supervisor multi-agent setup, what is the typical mechanism for delegating a subtask to a specialist agent?

Accepted Answer

Specialist agents are exposed as callable tools/handoffs the supervisor invokes

Answer

All agents share one mutable global prompt randomly

Answer

The supervisor fine-tunes each specialist at runtime

Answer

Delegation requires restarting the process per subtask

Question 30

What is the principle of least privilege as applied to agent tool design?

Accepted Answer

Grant each tool only the minimal capability/scope needed to limit damage

Answer

Give the agent admin access for flexibility

Answer

Expose one universal tool that does everything

Answer

Privilege only matters for the database layer

Test yourself