AI Agent interview questions, practical project and 30-day plan

A framework for answering design questions

1. Define the goal: who is the user and what counts as success?
2. Set boundaries: what belongs to the model, code or a human?
3. Split the chain: input, retrieval, planning, tools, state and output.
4. Describe data: context, memory, business data, vector data and permissions.
5. Give metrics: success, accuracy, latency, cost and security.
6. Explain failure: timeouts, permissions, conflicts, loops and model errors.
7. Explain evolution: evaluation, canary, rollback and tenant expansion.

Common questions and answer points

• What is an Agent? A model-centered task system that uses tools, state, memory, workflows and verification; the model proposes actions and the application authorizes and executes them.
• How are Agent and RAG related? RAG retrieves external knowledge; an Agent is the larger execution system that may call RAG and business APIs.
• When should you avoid an Agent? Use deterministic workflows for fixed, high-risk or strongly audited processes.
• How do you debug wrong RAG answers? Inspect parsing, index freshness, Recall@K, candidate relevance, filtering, reranking, context assembly, prompt constraints and citations.
• How do you prevent tool privilege escalation? Use allowlists, server-side schema and resource checks, least privilege, approval, idempotency and audit.
• Should every failed tool call be retried? Classify the error: argument and permission errors usually stop; timeouts may retry within bounds; conflicts require rereading state.

Practical project design

Build a company knowledge assistant that uses RAG for policy questions, read-only tools for status checks and a write tool that requires human approval before submission.

• Entry: sign-in, tenant, session and request ID.
• Knowledge: parsing, chunking, embeddings, hybrid retrieval and reranking.
• Orchestration: intent, retrieval, tool routing, state and stop conditions.
• Tools: read/write separation, schema, permissions and idempotency.
• Memory: session summary, user preference and task state.
• Evaluation: dataset, automated scoring, human review and Badcases.
• Operations: Docker, logs, Trace, alerts, canary and rollback.

Project walkthrough template

1. Business problem, current cost and user scale.
2. Target metrics such as success, first-token latency, handoff and unit cost.
3. Why the design uses RAG, workflow, single-agent or multi-agent execution.
4. Tool contracts, permissions, memory, recovery and evaluation.
5. Before/after evidence and the most serious remaining boundary.

Prepare one concrete failure story, such as an outdated policy being retrieved. Explain the diagnosis, version filter, index state and regression case rather than saying only “we improved the prompt.”

Live coding and troubleshooting

• Agent loop: messages, tool results, stop conditions, maximum steps and structured errors.
• Tool executor: allowlist, schema, timeout, retry, authorization and Trace.
• RAG retrieval: metadata filters, top-k, deduplication, ranking and source retention.
• Online timeout: split model, retrieval, tool, queue and downstream time with Trace.
• High concurrency: pools, queues, rate limits, caching, streaming, degradation and capacity.

30-day review plan

1. Days 1-3: classify roles and JD keywords; produce a gap list.
2. Days 4-7: build a minimal model API and structured-output demo.
3. Days 8-12: implement an Agent loop, state and workflow diagram.
4. Days 13-18: implement RAG ingestion, retrieval, citation and evaluation.
5. Days 19-21: add tool permissions, idempotency and MCP study notes.
6. Days 22-25: add Trace, metrics, automated evaluation and Badcase review.
7. Days 26-28: design capacity, cost, production architecture and runbook.
8. Days 29-30: rehearse a three-minute introduction, deep dive and live coding problem.

Final acceptance checklist

• I can define an Agent in sixty seconds and distinguish it from RAG and workflows.
• I can draw model, context, orchestration, tools, data, permissions and observability.
• I can explain RAG ingestion, retrieval, reranking, citations, updates and evaluation.
• I can implement and debug tool calling with timeout, permission, retry and idempotency.
• I can explain MCP Host, Client, Server, discovery and security boundaries.
• I have metrics for success, latency, cost and security.
• I can tell one failure story with diagnosis, fix, regression and release evidence.

Review the series in order: job map → foundations → RAG → tools and MCP → production.