Interview: Agent System Design

What you will be able to do

• ✓Use a repeatable framework for agent design questions
• ✓Design common agent systems end to end
• ✓Justify choices with trade-offs
• ✓Include safety, evaluation and cost in every design

The idea, in plain English

System design is where senior candidates are separated. There is no single right answer - the interviewer watches how you clarify, structure, choose and justify.

Use the same framework every time, so you never freeze: clarify the goal, list requirements, sketch the architecture, choose tools and data, define the control flow, add safety, plan evaluation, and discuss scale and cost.

The worked designs below reuse everything from this course: RAG, tools, LangGraph, MCP, multi-agent patterns, tracing and guardrails.

Worked example: A reusable design framework and six worked system designs, from support bot to autonomous actions.

Intermediate: What framework should I use for any agent design question?

Work through eight steps out loud: (1) clarify the goal and users, (2) requirements - functional, latency, cost, accuracy, risk, (3) high-level architecture, (4) tools and data sources, (5) control flow - workflow, single agent, or multi-agent, (6) safety - permissions, approvals, guardrails, (7) evaluation - metrics and test sets, (8) scale, cost and failure handling.

• •Ask questions first - interviewers reward clarifying.
• •Draw boxes and arrows as you talk.
• •End with risks and what you would build first.

Advanced: Design a customer support agent for an online shop.

Clarify: which channels, which questions (orders, refunds, product info), and what it may DO (look up orders, start a refund?). Architecture: a router classifies the request; FAQ and policy questions go to RAG over help articles; order questions use read-only order tools with the user's identity; refunds use a write tool with limits and human approval above a threshold; anything unclear hands off to a human agent with a summary.

• •Tools: get_order(order_id) scoped to the logged-in user; create_refund_request with amount limits.
• •Safety: injection filtering, no other users' data, approval for money.
• •Evaluate: resolution rate, correct handoffs, CSAT, cost per conversation.
• •Start with read-only answers; add actions later.

Advanced: Design a research assistant that writes reports with sources.

A planner breaks the topic into sub-questions; a researcher agent runs web or document search per sub-question and saves sources with quotes; a writer drafts only from saved sources; a checker verifies every claim has a source and flags gaps; a human reviews before publishing.

• •Agentic RAG with query rewriting and limits on searches.
• •Store sources in state so citations are exact.
• •Evaluate: citation accuracy, coverage, factual errors per report.
• •Cost control: cap searches and rounds per sub-question.

Advanced: Design a coding agent that fixes bugs in a repository.

Inputs: an issue and the repo. Tools: search code, read files, edit files, run tests - all inside a sandboxed container. Loop: reproduce the bug with a test, propose a change, run tests, iterate within a step limit, then open a pull request for human review - never push directly.

• •Sandbox with no secrets and limited network.
• •Context management: retrieve relevant files, not the whole repo.
• •Evaluate: tests passing, review acceptance rate, regressions.

Advanced: Design an agent that books travel and pays for it.

Split searching from acting: the agent searches flights and hotels with read-only tools and presents options; booking and payment are separate tools that require explicit user confirmation of the exact item and price, use idempotent requests, and run within spending limits.

• •Never let the model decide to pay on its own.
• •Handle partial failures: flight booked, hotel failed - compensating actions.
• •Audit log of every action; easy cancellation.

Advanced: Design an internal knowledge assistant for a 5,000-person company.

Ingest documents from wikis, drives and tickets through connectors (MCP servers per system), chunk and embed them with permissions stored as metadata, retrieve with hybrid search filtered by the user's access rights, answer with citations, and keep the index fresh with incremental sync.

• •Permissions at retrieval time are the hardest and most important part.
• •Evaluate per department; track unanswered questions to find missing docs.

Advanced: Design a multi-agent content pipeline (research, write, edit).

Start with a sequential pipeline - researcher, writer, editor - with clear expected outputs per task and a final human approval. Add a supervisor only if the order must change per request. Pass summaries between agents, not full histories.

• •Limits on rounds; a single owner of the final draft.
• •Compare cost and quality with a single-agent baseline.

Advanced: How do you estimate the cost of an agent system?

Estimate tokens per run: system prompt + tool descriptions + history + tool results, times the number of steps, plus output tokens - then multiply by price per token and runs per day. Add embedding and infrastructure costs.

• •Example: 8 steps x 3,000 input tokens = 24,000 tokens per run before output.
• •Then reduce: trim context, cache, cheaper models for easy steps.

Advanced: How do you decide between a workflow, a single agent and multiple agents?

Use a workflow when the steps are known; a single agent when the path depends on the input but one prompt and tool set can cover it; multiple agents only when roles, tools or context clearly do not fit in one - and you can afford the extra cost and complexity.

• •Say which one you chose and why - that is the answer they want.

Advanced: What would you build first, and how would you roll it out?

A narrow version for one use case with read-only tools, tracing and an evaluation set; test with internal users; add actions with approvals; expand to more users with monitoring and a rollback plan.

• •Ship value early; earn trust before adding autonomy.

Advanced: What are the biggest risks in your design, and how do you reduce them?

Name them concretely: wrong answers (evaluation, citations), harmful actions (approval, limits), data leaks (permissions, filtering), prompt injection (layers of defence), cost blow-ups (budgets), and outages (fallbacks).

• •Closing with risks shows senior judgement.

Advanced: How do you handle long-running agent tasks that take minutes or hours?

Run them as background jobs from a queue, save state with a checkpointer so they can resume, report progress to the user, allow cancellation, and notify when done.

• •Do not hold an HTTP request open for an hour.

Interview questions