Interview: Multi-Agent Systems

What you will be able to do

• ✓Decide when multiple agents are worth it
• ✓Explain the main coordination patterns
• ✓Explain CrewAI and AutoGen concepts
• ✓Design how agents share information
• ✓Name and prevent multi-agent failure modes

The idea, in plain English

Multi-agent systems are popular in demos and job descriptions, so interviewers check whether you know when they help - and when one good agent is better.

Expect pattern questions (supervisor vs hierarchical), framework questions (CrewAI roles and processes, AutoGen conversations), and failure-mode questions.

Lesson 3.8 built multi-agent graphs; Module 7 builds crews with sequential and hierarchical processes.

Worked example: Fourteen questions on multi-agent design, tied to LangGraph's multi-agent graphs (Lesson 3.8) and the CrewAI module.

Basic: What is a multi-agent system?

Several agents, each with its own role, instructions and tools, working together on a task - coordinated by a fixed process, a supervisor agent, or a conversation between them.

• •Like a team: researcher, writer, reviewer.
• •Each agent can be simpler than one agent doing everything.

Basic: Why use several agents instead of one?

When one prompt cannot hold all the instructions and tools, when tasks need different expertise or models, or when you want separate checking (one writes, one reviews).

• •Smaller, focused prompts are easier to get right.
• •But: more calls, more cost, more coordination.

Intermediate: What coordination patterns do you know?

Sequential pipeline: each agent's output feeds the next. Supervisor: a manager agent routes work to workers. Hierarchical: supervisors of supervisors. Hand-offs: an agent passes control to the right specialist. Group chat: agents talk in turns, with a manager choosing who speaks.

• •Sequential is the most predictable.
• •Supervisor and group chat are more flexible and less predictable.

Intermediate: What are the main concepts in CrewAI?

Agents (role, goal, backstory, tools), Tasks (description, expected output, the agent responsible), and a Crew that runs the tasks with a Process - sequential, or hierarchical with a manager.

• •The role and backstory become part of the system prompt.
• •Module 7 (Lessons 7.1 to 7.8).

Intermediate: How is AutoGen different from CrewAI?

AutoGen models agents as conversable participants that exchange messages - two-agent chats or a group chat with a manager. CrewAI models a team with defined roles and tasks run in a process.

• •AutoGen: conversation-first. CrewAI: task-first.
• •LangGraph: graph-first, most explicit control.

Intermediate: How do agents share information?

Through shared state (a graph's state object), by passing outputs as inputs to the next task, through a shared message history, or through a shared store like a database or memory.

• •Share only what the next agent needs - full histories cost tokens and add noise.

Intermediate: What is the sequential vs hierarchical process in CrewAI?

Sequential runs tasks in the given order, each with its assigned agent. Hierarchical adds a manager agent that decides which agent does what and checks the results.

• •Hierarchical is more flexible but slower, costlier and less predictable.
• •Lessons 7.5 and 7.6 compare them on the same task.

Advanced: What are common failure modes of multi-agent systems?

Agents passing work back and forth in loops, a supervisor that delegates badly, context lost between agents, duplicated work, unclear ownership of the final answer, and costs that multiply with every agent and round.

• •Set limits per agent and for the whole run.
• •Trace every message between agents.

Advanced: How do you control cost in a multi-agent system?

Limit rounds and delegations, use smaller models for simple roles, pass summaries instead of full histories, cache results, and measure tokens per agent in traces.

• •Token counts can be over-reported when agents share one model object (seen in Module 7) - check how your framework counts.

Advanced: How do you evaluate a multi-agent system?

Evaluate the end result on a task set, and also each agent's contribution: did the router pick the right agent, did each agent's output meet its task's expected output, how many rounds and tokens did it take?

• •Compare against a single-agent baseline - multi-agent must earn its cost.

Advanced: When is a multi-agent system the wrong choice?

When one agent with good tools can do the job, when the task is a fixed pipeline (use a workflow), or when latency and cost matter more than flexibility.

• •Many "multi-agent" demos work as well with one agent and three tools.

Advanced: How would you add a human into a multi-agent workflow?

Make the human a step in the process: pause before key decisions or final outputs, show what each agent produced, allow edits, then resume - with the state saved so nothing is lost while waiting.

• •LangGraph interrupts; AutoGen's user proxy agent; CrewAI's human input options.

Advanced: How do you prevent agents from endlessly delegating to each other?

Cap delegations and rounds, track who has handled a task, require each delegation to have a clear reason and target, and give a final owner who must answer when the limit is reached.

• •Disable delegation for agents that do not need it.

Advanced: Design a research-and-writing team of agents. What would you build?

A sequential pipeline: a researcher with search tools gathers sources, a writer drafts from those sources only, an editor checks facts against the sources and style, and a final check validates the format - with limits, tracing and a human review step before publishing.

• •Start sequential; add a supervisor only if the order must change per task.
• •Module 7's content crew project.

Interview questions