Multi-Agent Architectures: interview questions

Production questions of the kind asked in applied AI engineering interviews.

1. Larkspur's support swarm sometimes takes a minute to fail. The trace shows billing and operations handing the same customer back and forth until LangGraph raises GraphRecursionError. What do you look at, and what do you change?

What a strong answer covers:

• Reads the handoff ToolMessages and each agent's prompt: both agents believe the other owns "outage credits", so the ownership line is unclear
• Fixes ownership in the agents' prompts and handoff tool descriptions, and gives one of them the answer instead of another handoff
• Sets recursion_limit explicitly on every run (25 in the notebooks), because the default of 10,007 steps is what made the failure take a minute
• Adds a scripted-model test that replays the ping-pong and asserts the run ends with the error, so a prompt change cannot bring the loop back silently

2. The T9.3 router sends each question to one knowledge base. Customers ask "my server went down, do I get credits?" and get an answer about the outage with nothing about credits. Do you replace the router with a supervisor or a swarm?

What a strong answer covers:

• Names the failure: a one-shot router picks one specialist, so a two-part request loses its second half
• Supervisor: re-decides after each specialist, so it covers both halves; 5 calls for this request, with every decision in one log
• Swarm: operations answers its part and hands off to billing; 3 calls, and follow-ups go straight to the active agent
• Chooses by traffic shape: many multi-turn conversations with one specialist favour the swarm; audit or compliance needs favour the supervisor's single decision log

3. Your lab brief says to compare routing accuracy and cost after replacing the router with a supervisor over three specialist RAG agents. How do you measure it?

What a strong answer covers:

• Labels each golden question with the specialists it needs, in order, and scores the route the system took against that label
• Logs routing outside the transcript (route_log, one entry per decision with its reason), so routing accuracy is read from state and not parsed from prose
• Reports model calls, tokens and latency per request, split into routing calls and specialist calls, next to answer quality from the T5.1 baseline
• Looks per question, not only at the mean: single-domain questions get more expensive under a supervisor while multi-part ones get correct

4. A swarm thread has active_agent = billing. Three days later the same customer writes "Can I export my data?" What happens, and what should happen?

What a strong answer covers:

• The START router reads the saved active_agent, so billing receives a question it has no handbook section for
• Billing needs a handoff tool to the data agent and a prompt that tells it to use it; otherwise it answers from nothing or asks the customer to repeat
• The library handoff tool carries no note, so whatever data needs must already be in the messages; a hand-built tool can pass a context note
• Considers expiring or resetting active_agent after a quiet period, so a new topic on an old thread starts at triage

5. Larkspur wants a human to approve any refund the billing agent offers outside the 30-day window before the customer sees it. Where does the pause go in a supervisor workflow?

What a strong answer covers:

• Uses a checkpointer, because a paused run has to be saved and resumed by thread
• Prefers a dynamic interrupt(payload) inside the billing node, so only out-of-window refunds pause; a static interrupt_after pauses every run
• Lets the reviewer edit state (update_state) or resume with Command(resume=...), then sends the edited answer, not the model's
• Checks the window in code from the purchase date, so the decision to pause never rests on the model's judgement