LLM Engineer Interview Questions: LLM Agents, Tool Use, Multi-Step Reasoning, MCP Protocol

An LLM agent is a system where an LLM acts as a reasoning engine that plans, takes actions, and observes results in a loop. Unlike a single-turn chatbot, an agent can call tools, retrieve information, execute code, and refine its approach based on intermediate results. Agents enable LLMs to handle complex multi-step tasks autonomously.

An LLM agent typically has a planner that breaks the task into steps, a memory that stores intermediate state, a set of tools the agent can call, an executor that runs the tools, and a loop that continues until the task is complete or a stopping condition is met. Modern agent frameworks like LangGraph, AutoGen, and CrewAI provide these components as building blocks.

Tool use is the ability of an LLM to invoke external functions to accomplish tasks beyond pure text generation. Common tools include web search, code execution, database queries, calculators, and API calls. The LLM decides when to use a tool and produces structured arguments matching the tool's schema. Modern frontier models support tool use natively through their APIs.

The Model Context Protocol, or MCP, is an open standard introduced by Anthropic in late 2024 for connecting LLM applications to external data sources and tools. It defines a universal way for any client to talk to any server providing context, replacing custom integrations with a standardized protocol. MCP is widely adopted across the AI ecosystem in 2026.

Function calling is an API-level feature where you describe tools to a single LLM call. MCP is a protocol for separate servers that expose tools, resources, and prompts to any compatible LLM client. MCP enables a marketplace of reusable integrations: one MCP server for GitHub, one for Slack, one for a database, all usable by any MCP-aware client without custom code.

Multi-step reasoning is when an agent breaks a complex task into a sequence of smaller steps, executing each before moving on. Each step can involve thinking, using a tool, and observing results. This pattern enables agents to solve problems that no single LLM call could handle, like research tasks that require multiple searches and synthesis.

Sequential tool calling executes one tool at a time, using each result to inform the next call. Parallel tool calling, supported by Claude, GPT-4, and Gemini, lets the model invoke multiple independent tools in a single response. Parallel calling reduces latency when tasks have independent subtasks, like fetching weather for multiple cities at once.

Agent memory strategies include sliding windows that keep only the last N messages, summarization that compresses old turns into a summary, vector retrieval that embeds past turns and recalls relevant ones, and structured memory that stores facts as key-value pairs. Production agents often combine multiple strategies to balance context size, cost, and recall.

ReAct interleaves thinking and acting, deciding the next action based on the previous observation. Plan-and-Execute first generates a complete multi-step plan, then executes the steps in order. ReAct adapts better to surprises mid-task, while Plan-and-Execute is more efficient when the path is predictable. Many production systems combine both, planning at the high level and reacting within steps.

Common agent failures include infinite loops where the agent never terminates, hallucinated tool calls with wrong arguments, getting stuck on impossible subtasks, losing track of the original goal, and runaway costs from too many LLM calls. Production agents need step limits, cost ceilings, timeout handling, and clear stopping criteria.

A workflow has predefined steps and control flow, with LLM calls slotted into specific points. An agentic system gives the LLM autonomy to decide what to do next at each step. Workflows are more predictable, debuggable, and cheaper. Agents are more flexible but harder to control. Anthropic's research recommends starting with workflows and only adding agency where it provides clear value. ---