Collaboration, Control & Parallel Agents: interview questions

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

1. Larkspur's support assistant now answers each part of a compound question in a parallel branch. Since the change, some replies to "refund, export and outage credits" questions say nothing about credits, and the trace shows two synthesizer outputs for one request. What happened, and how do you fix it?

What a strong answer covers:

• Reads the trace by superstep: the outage branch has one more node (a query rewrite) than the others, so it finishes a superstep later
• With separate edges into the synthesizer, the join runs once per superstep in which a predecessor finished: first with three sub-answers, then again with four. The reply that went out was the first
• Fixes the join with a list edge (add_edge([...], "synth")) or defer=True, or keeps every branch one node long by moving the rewrite inside the agent
• Adds a deterministic test with stub nodes that counts synthesizer runs and the sub-answers each run saw, so the fix stays fixed

2. Escalations ask for one of two follow-ups depending on what the incident log shows: a service-credit reply if availability fell below the SLA, a status-page reply otherwise. Would you build this as a fixed pipeline, a router or a blackboard?

What a strong answer covers:

• A fixed pipeline cannot skip a step, so it runs both follow-ups on every escalation and pays for both
• A router decides once, before any specialist has run, so it cannot branch on what the incident analyst finds
• A blackboard controller re-reads the board after each post and can choose the credit branch only after the incident finding lands; that is the behaviour the request needs
• Guards it: a Literal decision so the controller can only name a real node, a call cap enforced in code, a recursion limit as a backstop, and a computed check of which branches ran

3. You parallelised the four sub-question agents. Median latency fell, but p95 barely moved and the provider started returning rate-limit errors. How do you measure what is going on, and what do you change?

What a strong answer covers:

• Wall time is the sum, over supersteps, of the slowest task in each, so p95 follows the slowest branch's tail; logs per-node latency from the updates stream to find it
• Caps concurrency (max_concurrency in the run config) to stay under the rate limit and measures what it costs: in the lab a cap of 2 takes the answer from 2.7 s to 3.7 s
• Tracks time to the first streamed sub-answer separately from time to the final answer, since that is what the user sees first
• Re-runs the golden set from the T5.1 baseline to confirm the faster path did not change answer quality

4. One of the four branches times out on a provider call. What happens to the three sub-answers that finished, and how should the assistant reply?

What a strong answer covers:

• A task error fails the whole superstep; the synthesizer never runs on that attempt
• With a checkpointer, LangGraph keeps the writes of the tasks that succeeded in that superstep, so a resume re-runs only the failed one
• Adds a retry policy to the flaky node, and decides whether a missing sub-answer should block the reply or become an explicit "could not check outage credits" line the synthesizer must state
• Alerts on the per-branch failure rate, since one flaky dependency now takes down every compound question that touches it

5. A product manager wants every refund answer to go through a three-round debate "so the assistant considers both sides". What do you say?

What a strong answer covers:

• A three-round debate costs 7 calls instead of 1; the notebook's own advice is that a factual question with one correct answer is better served by a single call
• Refund policy is a fact in the handbook; if two retrieved chunks disagree (an old and a new policy), the fix is version metadata and filtering at retrieval, not an argument
• Debate earns its cost on real trade-offs, such as whether to replace RAG with a long-context prompt, where the transcript and the judge's strongest point from each side are the deliverable
• Proposes a test: run debate and single-call answers on the golden set and compare correctness, faithfulness, cost and latency before deciding