LangChain Architecture Interview Questions

500M vectors at sub-100ms p99 is a staff-level vector search design question from Uber, Airbnb, and large-scale ML platform teams. Learn sharding strategies, index tuning, and the operational trade-offs that separate senior from principal engineers.

Cursor rules are context engineering in disguise.

A production AI coding agent typically includes:

**Session model** WebSocket/WebRTC streaming multimodal input/output.

**Proxy layer** Backend holds keys; assigns budgets per customer/plan.

**Session** Persistent realtime connection with session.update for instructions/tools; handle reconnect.

**Graph structure** draft → auto_checks → legal_subgraph → manager_subgraph → publish; reject edges loop to draft with comments in state.

**Supervisor pattern** Central node decides next worker or FINISH based on plan and worker outputs — alternative to flat handoff messages.

**State schema** ticket_id, customer_tier, messages, retrieved_docs, proposed_actions, sentiment, escalation_reason, resolution_code.

**Retry patterns** Tenacity-style retries inside node or dedicated retry wrapper node incrementing state.retry_count. Exponential backoff for transient MCP errors.

**Parallelism model** LangGraph can schedule independent nodes in same superstep concurrently — e.g., search three sources at once. Reducers combine outputs into unified state field.

**LangChain agents** Higher-level executors (ReAct, tool calling) with less boilerplate. Good for demos and simple loops. Harder to debug complex branching, human interrupts, and precise persistence semantics.

**Why subgraphs** Parent graph orchestrates; subgraphs own domain logic — billing dispute flow, code review pipeline. Teams ship subgraph versions independently.

**Cycle design** Classic ReAct loop: agent → tools → agent. LangGraph makes cycle explicit vs hidden while True loops. Conditional edge from tools checks whether to continue.

**Mechanics** After node N runs, a routing function reads state — e.g., last tool call, confidence score, intent label — and returns one or more next nodes. LangGraph supports static maps from route keys to targets. This replaces fragile if/else scattered in agent code.

Multi-server hosts are production reality — GitHub + Jira + DB + docs + internal APIs. Poor design creates tool soup.

Enterprise MCP rarely connects laptops directly to SaaS MCP servers. The gateway is the control plane interviewers want you to design.

Building a custom MCP server is a common system-design follow-up to "what is MCP?" Strong candidates describe end-to-end architecture, not just SDK boilerplate.

Horizontal scaling patterns: hash sharding, semantic partitions, routing embeddings, and merge/rerank at query time.

Protect instructions, secrets, policy, and tool policies in system prompts against leakage, override, and indirect injection.

Medium AI Agents interview question on tool schema design and versioning — architecture, trade-offs, eval, and production patterns.

Hard AI Agents interview question on customer support agent system design — architecture, trade-offs, eval, and production patterns.

Medium AI Agents interview question on swarm vs supervisor multi-agent design — architecture, trade-offs, eval, and production patterns.

Medium AI Agents interview question on idempotent tool execution — architecture, trade-offs, eval, and production patterns.