LangGraph Agents & Multi-Agent Systems interview questions

31 questions from the LangGraph Agents & Multi-Agent Systems course, each with the correct answer and an explanation. Try to answer before you open the answer.

Why Agents Need Graphs

1. What does a workflow graph offer over a plain chain?

  1. Free tokens
  2. A bigger model
  3. Branches, loops and pauses with explicit control flow
  4. No need for state
Show answer

Answer: Branches, loops and pauses with explicit control flow

Graphs express non-linear flows while keeping them inspectable.

Lesson: Chains, Agents and Workflows

2. What does a node return?

  1. A partial update to the state
  2. A new graph
  3. A trained model
  4. Nothing ever
Show answer

Answer: A partial update to the state

Nodes return only the keys they change; LangGraph merges them into the state.

Lesson: The LangGraph Model: State, Nodes, Edges

3. What does compile() return?

  1. A database
  2. A trained model
  3. A JSON file
  4. A runnable app with invoke and stream
Show answer

Answer: A runnable app with invoke and stream

Compiling validates the graph and produces the executable application.

Lesson: Your First Graph

4. What does a reducer on a state key do?

  1. Deletes the key
  2. Defines how new updates combine with the existing value
  3. Compiles the graph
  4. Speeds up the GPU
Show answer

Answer: Defines how new updates combine with the existing value

Without a reducer the latest write replaces the old value.

Lesson: State Reducers: Replace or Accumulate

Control Flow: Branches, Loops, Commands, Fan-Out

5. What should a router always cover?

  1. Only the first label
  2. A default path for unexpected labels
  3. Nothing
  4. A GPU
Show answer

Answer: A default path for unexpected labels

Unhandled labels cause errors or dead ends.

Lesson: Conditional Edges and Routing

6. What must every loop in a graph have?

  1. A bigger model
  2. An unlimited retry count
  3. No state
  4. A termination rule and a hard cap on iterations
Show answer

Answer: A termination rule and a hard cap on iterations

Caps prevent infinite loops and runaway cost.

Lesson: Loops: Retry, Refine and Reflect

7. What is the recursion limit for?

  1. Stopping runaway graphs as a safety net
  2. Making the model smarter
  3. Compressing state
  4. Choosing the language
Show answer

Answer: Stopping runaway graphs as a safety net

It bounds the number of steps in a single run.

Lesson: Recursion Limit: The Safety Net

8. What does returning Command(update=..., goto=...) do?

  1. Calls the model twice
  2. Deletes the graph
  3. Updates the state and chooses the next node in one step
  4. Sets the recursion limit
Show answer

Answer: Updates the state and chooses the next node in one step

It combines a state write with a routing decision.

Lesson: Command: Update State and Route Together

9. When is Send most useful?

  1. To disable reducers
  2. For a fixed two-step chain
  3. When the number of parallel tasks is known only at run time
  4. To compile the graph
Show answer

Answer: When the number of parallel tasks is known only at run time

Send creates one task per item dynamically.

Lesson: Fan-Out With Send (Map-Reduce)

Agents With Tools

10. When does the agent loop end?

  1. When the model answers without requesting a tool, or a limit is hit
  2. After exactly one tool call
  3. Never
  4. When the GPU is idle
Show answer

Answer: When the model answers without requesting a tool, or a limit is hit

The conditional edge routes to END once no tool is requested.

Lesson: The Tool-Calling Agent as a Graph

11. When should you build your own graph instead of using a prebuilt agent?

  1. When you have no tools
  2. Never
  3. Only for one-line scripts
  4. When you need custom routing, approvals or extra nodes
Show answer

Answer: When you need custom routing, approvals or extra nodes

Custom graphs give full control over the flow.

Lesson: Prebuilt Agents With a Real Model

12. Which tool design is better for an agent?

  1. One huge tool that does everything
  2. One narrow, well-named tool with a clear description and recoverable errors
  3. Tools with no descriptions
  4. Tools that return raw database dumps
Show answer

Answer: One narrow, well-named tool with a clear description and recoverable errors

The model relies on names, descriptions and error messages to use tools well.

Lesson: Designing Tools That Agents Use Well

13. Why does agent cost grow faster than the number of steps?

  1. Tokens get cheaper with time
  2. Each step re-reads a longer message history
  3. Tools are free
  4. Graphs forbid caching
Show answer

Answer: Each step re-reads a longer message history

Growing context makes later calls more expensive than earlier ones.

Lesson: Agent Limits, Costs and Stop Conditions

Memory, Persistence and Time Travel

14. What identifies a conversation to the checkpointer?

  1. The GPU id
  2. The model name
  3. The node name
  4. The thread_id in the run config
Show answer

Answer: The thread_id in the run config

Each thread_id has its own saved state history.

Lesson: Checkpointers and Threads

15. What does time travel let you do?

  1. Inspect and resume from earlier checkpoints of a thread
  2. Change the past of the real world
  3. Skip all nodes
  4. Disable checkpoints
Show answer

Answer: Inspect and resume from earlier checkpoints of a thread

Each checkpoint is a restorable snapshot of the run.

Lesson: State History, Replay and Time Travel

16. Which belongs in long-term memory?

  1. A node's local variable
  2. The intermediate tool result of one step
  3. A user's preferred language across conversations
  4. The recursion counter
Show answer

Answer: A user's preferred language across conversations

Cross-conversation facts live in the store; per-run details live in thread state.

Lesson: Long-Term Memory With a Store

17. Why summarise old messages?

  1. To disable tools
  2. To delete the thread
  3. To keep context short and cheap while preserving key information
  4. To increase the recursion limit
Show answer

Answer: To keep context short and cheap while preserving key information

Summaries compress history so it fits the window.

Lesson: Managing Long Histories: Trim and Summarise

Human Approval, Reliability and Streaming

18. What does interrupt() require to work?

  1. A checkpointer and a thread_id
  2. A GPU
  3. A second graph
  4. A recursion limit of one
Show answer

Answer: A checkpointer and a thread_id

The paused state must be saved so the run can be resumed later.

Lesson: Human-in-the-Loop With interrupt()

19. For a permanent failure, what is a good pattern?

  1. Increase the temperature
  2. Retry forever
  3. Ignore it silently
  4. Record an error in state and route to a fallback or human
Show answer

Answer: Record an error in state and route to a fallback or human

Explicit error state makes failures handleable by the graph.

Lesson: Retry Policies and Error Handling

20. Which stream mode emits only what each node changed?

  1. values
  2. updates
  3. none
  4. checkpoint
Show answer

Answer: updates

"updates" gives per-node deltas; "values" gives the full state each step.

Lesson: Streaming Updates, Values and Tokens

Multi-Agent Systems

21. What is a real cost of multi-agent systems?

  1. They are always cheaper
  2. More calls, latency and coordination bugs
  3. They remove the need for testing
  4. They need no prompts
Show answer

Answer: More calls, latency and coordination bugs

More moving parts mean more cost and more ways to fail.

Lesson: Why and When to Use Multiple Agents

22. In the supervisor pattern, who keeps overall control flow?

  1. Nobody
  2. Each worker independently
  3. The user's browser
  4. The supervisor
Show answer

Answer: The supervisor

Workers report back; the supervisor decides what happens next.

Lesson: The Supervisor Pattern

23. What is a typical risk of hand-off networks?

  1. Agents ping-ponging control back and forth
  2. Too few model calls
  3. No need for state
  4. Automatic correctness
Show answer

Answer: Agents ping-ponging control back and forth

Without hop limits, agents can pass a task around forever.

Lesson: Hand-Offs and Swarm-Style Networks

24. What is a benefit of sub-graphs?

  1. They are always faster
  2. They remove all state
  3. Reusable, independently testable components
  4. They replace tools
Show answer

Answer: Reusable, independently testable components

A compiled graph can be tested alone and plugged in as a node.

Lesson: Sub-Graphs and Hierarchical Teams

Testing, Evaluating and Deploying

25. How do you test a whole graph without calling a real LLM?

  1. Call production for every test
  2. You cannot
  3. Replace the model node with a scripted fake
  4. Delete the graph
Show answer

Answer: Replace the model node with a scripted fake

A deterministic fake makes tests fast, free and repeatable.

Lesson: Testing Graphs Without a Model

26. Why evaluate the trajectory and not just the final answer?

  1. An agent can be right for the wrong or unsafe reasons
  2. Trajectories are shorter
  3. Final answers are never useful
  4. It avoids tracing
Show answer

Answer: An agent can be right for the wrong or unsafe reasons

The path reveals wasted steps, wrong tools and policy violations.

Lesson: Evaluating Agents: Trajectories, Tools and Outcomes

27. Where should the thread_id come from in production?

  1. The model name
  2. Chosen by the client freely
  3. Always the same constant
  4. Derived from the authenticated user and session
Show answer

Answer: Derived from the authenticated user and session

Otherwise one user could load another user's conversation.

Lesson: Deploying, Observability and Guardrails

Putting It Together

28. Why is the refund step behind an interrupt?

  1. Interrupts make tools faster
  2. It is costly and irreversible, so a human should approve it
  3. Refunds are free
  4. To avoid state
Show answer

Answer: It is costly and irreversible, so a human should approve it

Human approval limits the damage of a wrong or manipulated action.

Lesson: Case Study: A Customer-Support Agent Team

29. Two parallel nodes both write to the same state key in one step. What do you need?

  1. A bigger model
  2. A reducer that defines how to merge the updates
  3. A new checkpointer
  4. Nothing, it always works
Show answer

Answer: A reducer that defines how to merge the updates

Without a reducer, concurrent writes to one key conflict.

Lesson: Revision: Cheat Sheet and Self-Check

30. Your agent sometimes loops forever calling the same tool. Which combination helps most?

  1. Removing the checkpointer
  2. A higher temperature
  3. More tools
  4. A step cap in state, a recursion limit, and a clear stop condition
Show answer

Answer: A step cap in state, a recursion limit, and a clear stop condition

Layered limits bound the loop even when the model never stops.

Lesson: Revision: Cheat Sheet and Self-Check

31. What lets a paused interrupt survive a server restart?

  1. A durable checkpointer storing the thread state
  2. Keeping the process in memory only
  3. A larger context window
  4. A faster GPU
Show answer

Answer: A durable checkpointer storing the thread state

The saved checkpoint is what is resumed later, even by another process.

Lesson: Revision: Cheat Sheet and Self-Check

← Back to the course