Agentic AI Interview Questions — 20 Top Q&As (2026)

Foundations & Patterns (5)

→Q: ReAct vs Plan-and-Execute? A: ReAct is reactive (think-act each step); Plan-and-Execute plans once then runs. ReAct is robust to failure; Plan-Execute is cheaper.
→Q: When do you NOT use an agent? A: When the task is fixed (pipeline wins) or facts-lookup (RAG wins). Agents add variance; only use when you need autonomy.
→Q: Tree of Thought — when worth it? A: Hard reasoning tasks where wrong intermediate steps are catastrophic; expensive — use only when accuracy beats cost.
→Q: How do you handle tool errors? A: Retry with backoff for transient; reflect-and-retry for semantic errors; escalate to human for irreversible.
→Q: How does function calling actually work? A: LLM emits a JSON object matching the tool schema; your runtime executes it and feeds the result back as a tool message.

→Q: Multi-agent — when and how? A: Use for parallelism or specialization. Supervisor (router), Swarm (peer), Graph (custom), Hierarchical (team-of-teams).
→Q: How do you prevent context explosion in multi-agent? A: Each agent gets only its slice; summarize handoffs; use blackboard pattern for shared state.
→Q: Short-term vs long-term memory? A: Short = conversation buffer with summarization; long = vector store; MemGPT for agent-managed hierarchical memory.
→Q: How does CrewAI differ from LangGraph? A: CrewAI is role-based (Crew/Agent/Task); LangGraph is graph-based with explicit state. Both work; pick by mental model.
→Q: How do you eval a multi-agent system? A: Per-agent trajectory eval + end-to-end task success + cost-per-task + human review of edge cases.

→Q: What is MCP? A: Model Context Protocol — Anthropic's open standard for tool servers; any MCP-compatible client (Claude Desktop, IDEs) can use any MCP server.
→Q: What is A2A? A: Google's Agent-to-Agent protocol; lets agents from different vendors discover and call each other; cross-platform interop.
→Q: OpenAI Agents SDK — what's unique? A: Handoffs between agents, built-in guardrails, native tracing; OpenAI's official agent framework (2025+).
→Q: Where would you NOT use LangChain? A: Hot inference paths where its abstractions add latency; deeply custom agents where the abstraction fights you.
→Q: How do you compare agent frameworks? A: State management (explicit vs implicit), observability hooks, retry semantics, multi-agent support, vendor lock-in.

→Q: How do you make an agent safe for production? A: Tool allowlist, budget limits, irreversible-action approvals, human-in-the-loop, audit logs.
→Q: What goes in agent observability? A: Trace every LLM call, tool call, retry; correlation IDs; cost per request; LangSmith / OpenTelemetry spans.
→Q: How do you eval an open-ended agent task? A: LLM-as-judge with structured rubric, sampled human review, regression suite on past examples.
→Q: How do you debug a hallucinated tool call? A: Inspect the trace, check prompt for ambiguity, verify schema, add few-shot examples, validate before execution.
→Q: Walk me through deploying an agent. A: Containerize, serverless or persistent worker, message queue for long-running tasks, retries with dead-letter queue, observability from day one.