Four Core Topics to Master for Big‑Tech Agent Interviews

1. Basic Understanding – What Is an Agent?

Interviewers often ask whether a project uses a plain LLM or an Agent and why. The key differences are:

LLM : a conditional‑probability model that takes a prompt and returns text; stateless, no memory, only generates language.

Agent : combines LLM, tools, memory, and planning to autonomously achieve goals; has short‑term and long‑term memory and can invoke external tools.

One‑sentence summary: LLM tells you how to act, an Agent does it for you.

Agent vs. Workflow

Comparison of control, token consumption, flexibility, and suitable tasks:

Controller : Workflow is coded by developers, Agent is driven by the LLM.

Token cost : Workflow ~1×, Agent ~4–8×.

Flexibility : Workflow low, Agent high.

Task suitability : Workflow fits fixed processes; Agent excels at open‑ended goals.

Interview bonus: the most common production architecture mixes Workflow (stable skeleton) with Agent (handling exceptions and complex cases).

Four‑Module Agent Architecture

LLM (brain) : understands intent and performs reasoning.

Planning module : decomposes tasks and orders steps.

Memory module : short‑term context + long‑term knowledge store.

Tool module : calls external APIs, databases, or code executors.

2. Execution Modes – How Does an Agent Run?

ReAct (Reason → Act → Observe)

The classic loop: Thought → Action → Observation → back to Thought. To avoid infinite loops, interviewers expect three safeguards:

Maximum step limit (commonly 15 steps).

Duplicate‑action detection – three consecutive identical tool‑calls with the same parameters cause termination.

Timeout control – set a maximum execution time for the task.

Plan‑and‑Execute

Generates a complete plan once, then executes step by step, consuming roughly 20 % of the tokens that ReAct uses.

Reflection

One Agent generates output, a second Agent reviews it; the loop repeats until quality criteria are met. Useful for code generation, legal documents, and other high‑quality‑requirement scenarios.

Multi‑Agent Collaboration

An orchestrator assigns tasks to specialized agents such as Research, Coder, and Reviewer. Anthropic cautions against premature use of many agents: a single strong Agent is often more stable and cheaper than several weak ones.

3. Tools & Protocols – Connecting an Agent to the Outside World

Function Call

The LLM only outputs the name of a function and its arguments; the actual execution is performed by the host application, which improves safety.

GPT‑4o and Claude 3.5+ support Parallel Function Call, reducing latency from T₁+T₂+T₃ to max(T₁,T₂,T₃).

Model Context Protocol (MCP)

Solves the N × M integration explosion: instead of writing separate integration code for each of 3 applications × 3 tools (9 codebases), MCP reduces it to 3 + 3 = 6.

MCP Host : the AI application (e.g., Claude Desktop, Cursor).

MCP Client : the translator that talks to the server.

MCP Server : exposes concrete tool capabilities.

Resources: Tools, Data, Prompt templates.

A2A Protocol

Addresses Agent‑to‑Agent communication, complementing MCP which handles Agent‑to‑Tool calls.

Skills

Encodes domain‑expert knowledge as a package containing identity, workflow, cautions, and output format. The chain is: Skills define "how to think" → MCP defines "what to use" → Function Call defines "how to invoke".

4. Engineering Implementation – Turning an Agent into a Product

Memory System Design

A three‑level storage architecture:

Working memory : runtime variables/prompt fields, < 1 KB.

Short‑term memory : current session history stored in Redis (TTL ≈ 30 min, recent 20 turns).

Long‑term memory : user profiles and cross‑session preferences stored in a vector database + PostgreSQL, effectively unlimited.

Key operations: write, read, compress (summarise when token limits are hit), and expire (LRU + confidence decay).

Security & Reliability

Four major threats:

Prompt Injection – malicious commands injected via web content.

Privilege Escalation – Agent coerced into unauthorized actions.

Data Leakage – sensitive information sent to third parties.

Resource Abuse – endless loops causing API overuse.

Core defenses: least‑privilege principle and Human‑in‑the‑Loop for critical operations such as data deletion, external email sending, or permission changes.

Cost Optimisation (Token Consumption)

Tool selection – provide only truly needed tools to the Agent.

Mode selection – use Workflow for simple tasks (saves ~4× tokens).

Context compression – summarise historical dialogue.

Model routing – small models for simple tasks, large models for complex reasoning.

Caching – cache tool results and prompts.

Production Pitfalls

Five common issues and mitigations:

Infinite loops – enforce max steps and duplicate‑action detection.

Hallucinated tool calls – strictly validate tool names.

Context contamination – truncate sensibly and reset tasks when needed.

Token explosion – truncate tool outputs and paginate.

Prompt injection – separate data from commands.

Conclusion

The interview logic follows four lines: (1) basic concepts – what an Agent is and how it differs from LLM/Workflow; (2) execution modes – when to use ReAct, Plan‑and‑Execute, Reflection, or Multi‑Agent; (3) tool & protocol stack – Function Call → MCP → A2A → Skills; (4) engineering – memory design, security, cost control, and real‑world pitfalls. Explaining these points in order—from "what" to "why" to "how"—yields a much higher interview score than rote memorisation.