Question 1 of 40
What is an AI agent in product terms?
Answer options for question 1 A. Any static if/else script with no model involved in decisionsB. An LLM that can observe, decide, use tools, and act in a loop toward a goalC. A vector database index that stores embeddings without generationD. A single one-shot completion with no tool access or state
Question 2 of 40
Workflow vs agent — the key difference:
Answer options for question 2 A. Agents cannot call external APIs or read from databasesB. They are identical — both always use the same fixed pipelineC. Workflows follow fixed steps; agents choose steps dynamically from observationsD. Workflows always use LLMs; agents never call any language model
Question 3 of 40
Tool calling (function calling) lets the model:
Answer options for question 3 A. Only read local PDF files without returning results to contextB. Retrain its weights on every API response before continuingC. Skip the need for any schema or argument validation entirelyD. Emit structured requests to run code/APIs, then continue with tool results
Question 4 of 40
Why define a JSON schema for each tool?
Answer options for question 4 A. So the model knows valid function names and argument shapesB. To replace the LLM entirely with a traditional rules engineC. Schemas are optional decoration with no effect on tool callsD. To increase hallucinations by giving the model more freedom
Question 5 of 40
Planning vs execution in agent design:
Answer options for question 5 A. Planning means fine-tuning; execution means RAG retrieval onlyB. Both roles must be merged into one single undivided promptC. Planner proposes steps; executor runs tools/APIs and returns factsD. Execution always happens before any plan is written or reviewed
Question 6 of 40
A ReAct-style loop alternates between:
Answer options for question 6 A. User authentication and session logout on every iterationB. Training on new data and running inference on the same batchC. Encoding and decoding images without any text reasoningD. Reasoning (thought) and acting (tool use) until the task is done
Question 7 of 40
Why do agents need memory ?
Answer options for question 7 A. Context windows are limited; preferences and past steps must persistB. Memory removes the need for tools and external API callsC. LLMs already remember every user forever by default at inferenceD. Memory only stores raw HTTP headers from each API request
Question 8 of 40
Short-term vs long-term agent memory:
Answer options for question 8 A. Long-term memory is the same as the model's pretraining weightsB. Short-term = current session context; long-term = persisted profile across sessionsC. Short-term never expires and holds all data from every userD. Short-term lives on GPU; long-term is always public to all users
Question 9 of 40
Multi-agent systems are useful when:
Answer options for question 9 A. Tasks are trivial one-liners that need no planning or toolsB. You want maximum cost and latency on every simple requestC. You have no tools or APIs and only need plain text repliesD. Specialized roles work better than one giant prompt for complex tasks
Question 10 of 40
LangGraph compared to a single LangChain chain is often chosen for:
Answer options for question 10 A. Cycles, branching, and explicit agent state machinesB. Replacing Python with SQL as the only orchestration languageC. Training transformer models from scratch on custom corporaD. Removing all observability and logging from agent pipelines
Question 11 of 40
Before running a tool that books a flight , best practice is:
Answer options for question 11 A. Disable all logging for speed and skip argument validationB. Validate args, confirm with user for irreversible actions, log the callC. Let the model pick random cities if the user input is unclearD. Execute silently with no confirmation to reduce latency
Question 12 of 40
Retry / failure handling in agents should:
Answer options for question 12 A. Crash the server without returning any message to the userB. Hide all errors from the model so it continues blindlyC. Catch tool errors, surface them to the model, and retry or replan with limitsD. Retry infinitely with no backoff until the API succeeds
Question 13 of 40
Showing reasoning steps to users helps because:
Answer options for question 13 A. It builds trust and makes debugging agent mistakes easierB. It guarantees factual correctness on every tool resultC. It replaces the need for tools and external data sourcesD. It hides API failures so users never see error states
Question 14 of 40
LangChain at a high level provides:
Answer options for question 14 A. A full replacement for PyTorch when training custom modelsB. Connectors for models, tools, retrievers, and composable LLM pipelinesC. Low-level drivers for camera hardware and image captureD. Only CSS styling components for building chat user interfaces
Question 15 of 40
In a travel planner , an executor agent primarily:
Answer options for question 15 A. Only stores embeddings with no access to live external APIsB. Writes creative poetry unrelated to the user's travel requestC. Fine-tunes the base model weights on every itinerary changeD. Calls weather/maps APIs and returns structured facts to the planner
Question 16 of 40
A chatbot vs an agent — the main product difference:
Answer options for question 16 A. Chatbots always use tools; agents never call any external APIB. Agents cannot use LLMs and rely only on hardcoded rulesC. Agent loops with tools/actions toward a goal; chatbot often one-shot replyD. They are identical architectures with the same control flow
Question 17 of 40
After your app executes a tool , you should return results to the model as:
Answer options for question 17 A. Nothing — the model should guess what the tool returnedB. A structured tool result message in the conversationC. Only binary images with no text the model can readD. Raw API keys and secrets so the model can debug itself
Question 18 of 40
Running a tool that charges a credit card without user confirmation is risky because:
Answer options for question 18 A. Agents cannot call external APIs for payment processingB. Tools never receive arguments from the model in practiceC. JSON schemas prevent all errors so confirmation is unnecessaryD. Models can call tools with wrong args — destructive actions need approval
Question 19 of 40
ReAct (Reason + Act) pattern alternates:
Answer options for question 19 A. Only GPU training phases with no inference at runtimeB. CNN pooling and flatten layers without any language modelC. Thought/reasoning steps with tool actions and observationsD. Deleting all memory after each step to save context tokens
Question 20 of 40
A scratchpad in agent memory holds:
Answer options for question 20 A. Intermediate notes and plan for the current task sessionB. The entire internet indexed for retrieval on every turnC. Only CNN convolution filter weights from an image modelD. Permanent user passwords stored across all future sessions
Question 21 of 40
Storing “user prefers window seats” in MongoDB for future trips is:
Answer options for question 21 A. A replacement for tool calling and external API accessB. Short-term chat buffer only — lost when the session endsC. Long-term memory — injected when relevant in later sessionsD. The same as updating the model's pretraining corpus
Question 22 of 40
Capping max agent iterations (e.g. 10 steps) prevents:
Answer options for question 22 A. Runaway tool loops and unbounded API costB. All factual errors in model outputs and tool resultsC. Users from reading the UI during long agent runsD. Memory from working across multiple conversation turns
Question 23 of 40
Handoff between specialized agents means:
Answer options for question 23 A. Deleting the planner agent so only one model call remainsB. Only one LLM call allowed per application lifetimeC. All agents share one identical prompt and tool set foreverD. One agent passes state/task to another with a different role
Question 24 of 40
LangGraph is often chosen when you need:
Answer options for question 24 A. Removing all tools so the model answers from memory onlyB. Explicit graph/state-machine control over agent steps and branchesC. Only static HTML pages with no dynamic orchestration logicD. Training CNNs from scratch on a custom image dataset
Question 25 of 40
When a weather API fails , good agent design:
Answer options for question 25 A. Retries with backoff, surfaces error to planner/user, may replan without weatherB. Invents random temperatures silently so the user sees no delayC. Crashes with no message and leaves the UI in a loading stateD. Ignores the failure and books flights using guessed weather data
Question 26 of 40
A DAG in workflow design is:
Answer options for question 26 A. A random number generator for LLM sampling temperatureB. The same as an agent that chooses tools dynamically each turnC. A fixed flowchart of steps (A→B→C) with no loops back — engineer picks the pathD. A type of GPU memory used only during transformer training
Question 27 of 40
Model Context Protocol (MCP) standardizes:
Answer options for question 27 A. How to delete all agent memory on every tool callB. How to train transformer weights from scratch on one laptopC. How browsers render CSS for chat user interfaces onlyD. How apps expose tools and data sources to LLM clients in a reusable way
Question 28 of 40
Context engineering means:
Answer options for question 28 A. Writing one static system prompt and never changing it againB. Deliberately assembling system, memory, retrieval, and tool results each agent turnC. Dumping the entire chat history unbounded into every LLM callD. Removing all tools so the model answers from memory only
Question 29 of 40
LLM-as-judge in evals means:
Answer options for question 29 A. A second model scores outputs against a rubric (e.g. faithfulness 1–5)B. Replacing all human review forever with zero spot checksC. Only measuring GPU temperature during inferenceD. The production model grades its own homework with no rubric
Question 30 of 40
Trajectory eval for agents checks:
Answer options for question 30 A. Whether users clicked the logo in the navigation barB. Only the font size used in the chat UIC. Whether the model updated its pretraining weights each stepD. Whether the right tools ran in a sensible order — not just final text quality
Question 31 of 40
An idempotent tool call means:
Answer options for question 31 A. The tool must fail on the second attempt alwaysB. The model never receives tool results in contextC. Retrying with the same idempotency key does not create duplicate side effectsD. Only read-only tools are allowed in every agent system
Question 32 of 40
Stateless API servers for agents mean:
Answer options for question 32 A. Users cannot have multi-turn conversationsB. Session state lives in Redis/DB — any replica can handle the next requestC. Tools cannot be called from horizontally scaled deploymentsD. The server stores all chats in RAM and never restarts
Question 33 of 40
Guardrails in agent systems should be:
Answer options for question 33 A. Layered — input, tool allowlists, output validation, human approval for risky actionsB. Removed to maximize token throughputC. Optional — models never make mistakes on tool argumentsD. Applied only after users complain in production
Question 34 of 40
When should you NOT use an agent?
Answer options for question 34 A. User goal needs live API data that changes each hourB. Steps depend on intermediate tool observationsC. Task requires multiple tool calls based on prior resultsD. Single-doc RAG Q&A with good retrieval — one LLM call is enough
Question 35 of 40
Independent tool calls (weather for Paris and Lyon) can often:
Answer options for question 35 A. Require fine-tuning the base model between each callB. Replace the need for argument validation entirelyC. Run in parallel to reduce latency when there is no dependencyD. Never run in the same agent session
Question 36 of 40
When agent context nears the token limit, best practice is:
Answer options for question 36 A. Delete the system prompt entirelyB. Summarize older tool traces and keep scratchpad + recent turnsC. Paste more raw JSON from APIs to improve qualityD. Disable all guardrails to save tokens
Question 37 of 40
A supervisor agent pattern:
Answer options for question 37 A. Routes tasks to specialist workers and may review their outputB. Means only one LLM call is allowed for the entire applicationC. Forbids any tools or external API accessD. Is identical to a single prompt with 50 tools listed
Question 38 of 40
Mocked tools in CI evals are used to:
Answer options for question 38 A. Hide all agent traces from developers permanentlyB. Replace the need for any eval datasetC. Train the production model on real credit card numbersD. Return deterministic fixtures so tests are stable without live APIs
Question 39 of 40
Structured trace logs for agents should include:
Answer options for question 39 A. Raw API keys returned from environment variablesB. trace_id, tool names, latency, token usage, and success/failure per stepC. Only the final answer text with no intermediate stepsD. User passwords so support can debug faster
Question 40 of 40
Interview: RAG chatbot vs tool-using agent for “What is our refund policy?”
Answer options for question 40 A. Agent always required even for static FAQ textB. Neither — fine-tune the model weekly on every policy edit onlyC. RAG — retrieve policy doc, one grounded answer; agent is overkillD. Skip retrieval and let the model guess from pretraining