Agentic Workflow Patterns: interview questions

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

1. Larkspur's reply drafter is an evaluator-optimizer loop. After an evaluator model upgrade, p95 latency triples and cost per ticket doubles, but no request fails. What happened, and what do you change?

What a strong answer covers:

• Suspects the loop first: the new evaluator approves less often, so more tickets run to the iteration cap, and the cap returns a draft instead of an error
• Confirms it from logs: drafts per ticket, the share of runs ending at iteration_cap_reached, and which criterion fails most
• Fixes the criteria, not the cap: one boolean per criterion, counts and lengths checked in code, and a re-check of the new evaluator on drafts already graded by hand
• Alerts on the cap-hit rate as its own metric, and marks capped answers so they are not sent as approved

2. Larkspur routes support questions by model tier. "Can I get a refund after 45 days?" needs proration arithmetic, and the triage step sends it to the cheap tier, which answers it wrong with confidence. Which way should the router err, and how do you prove the router pays for itself?

What a strong answer covers:

• Names the asymmetry: over-escalation costs money and shows on a bill; under-escalation returns a wrong answer nobody sees until a customer acts on it
• Biases triage toward hard on borderline cases and escalates before answering, since detecting a confident wrong answer afterwards is the hard problem
• Measures routed accuracy and reasoning-call share against all-cheap and all-reasoning baselines on a labelled, honestly mixed set
• Reads the router's own confusion matrix, counting hard-sent-cheap separately, and remembers that triage is one more call on every request

3. The assistant fans out to the handbook index, the status page and the account API, then merges. p95 latency is 9 s and nobody knows which branch is slow. How do you find out, and what can you change?

What a strong answer covers:

• Times each node as it finishes (stream_mode="updates" in LangGraph, or per-run traces), since the fan-in waits for the slowest branch and the mean hides it
• Gives each branch a timeout and a fallback value, so one slow source degrades the answer instead of stalling it
• Checks the concurrency cap against the API rate limits and submits the slowest branch first when a cap applies
• Checks whether the merge step itself dominates: it needs every branch, so no fan-out can shorten it

4. You rebuild query decomposition as orchestrator-worker. For "How long does a refund take?" the planner returns five sub-questions, including "Introduction" and "Conclusion". What went wrong, and what guards do you add?

What a strong answer covers:

• Spots padding: a prompt with numeric anchors ("simple topics need 2, broad topics 5+") produced five sections for every topic in lesson 4 of T7.3
• Rewrites the planner prompt to ask for one item per distinct question and forbids padding sections
• Bounds the plan in code: a maximum number of Sends, a direct path for single-part questions, and a defined result for an empty plan
• Checks that workers consume their item, by confirming the set of workers that ran changes with the plan

5. A product manager wants to replace the fixed refund workflow with a ReAct agent "so it can handle anything". When do you agree, and what do you keep?

What a strong answer covers:

• Keeps a workflow where the steps are known: its calls are countable, each node and router can be unit-tested, and it stops at a known node with a reason
• Uses an agent where the steps depend on what the model finds, such as a multi-step account investigation, and accepts trajectory-level evaluation and a step limit as the bound
• Proposes a hybrid: a router that sends known intents to workflows and only the rest to the agent
• Compares both on the same golden questions for accuracy, LLM calls, p95 latency and the share of runs that hit a limit