Question 1

What is an AI agent, and how does it differ from a simple LLM call?

Accepted Answer

An AI agent is an autonomous system that uses an LLM as its reasoning engine to determine a sequence of actions, utilize external tools, and maintain state to achieve a complex goal. See video: AI Engineering Explained: LLM, RAG, MCP, Agent, Fine-Tuning, Quantization and AI Agent Explained. A simple LLM call is reactive and stateless—you send a prompt, it generates text, and it stops. An AI agent is proactive and stateful. It runs in a loop: it observes an input, thinks about what to do, acts by calling a tool (like a web search or a Python script), observes the tool's output, and repeats this cycle until the overarching goal is completed.

Question 2

What are the different types of AI Agent Memory?

Accepted Answer

Agent memory is divided into short-term memory (the immediate context window of the current conversation) and long-term memory (external storage, like a vector database, used to recall past interactions across sessions). See: AI Agent Memory. Because LLMs are inherently stateless, agents must artificially construct memory. Short-term memory is managed by appending the chat history to the API payload. Long-term memory involves saving key facts or interactions to a database, and then retrieving them via semantic search when the agent recognizes a relevant context in a future session. Episodic memory refers to remembering specific past events or actions.

Question 3

What is Harness Engineering in AI?

Accepted Answer

Harness Engineering involves building the robust infrastructure, sandboxes, and safety wrappers that allow AI agents to execute code or tools securely and reliably without crashing the host system. See: Harness Engineering in AI. When an agent writes and executes code to solve a problem (e.g., a data analysis agent running a Python script), it needs a safe execution environment. Harness engineering focuses on building secure Docker containers, setting up API gateways, defining strict IAM permissions for the agent, and handling stdout/stderr streams to feed errors back to the agent for self-correction.

Question 4

Explain the ReAct (Reasoning + Acting) agent architecture.

Accepted Answer

ReAct is a foundational agent architecture that forces the LLM to explicitly generate reasoning traces (Thought) alongside executable actions (Action), allowing it to dynamically plan, execute, and observe. See: ReAct Agent. ReAct interleaves reasoning and acting. By forcing the model to explicitly state its 'Thought' before taking an 'Action', the model can better track its progress, correct mistakes, and plan its next step based on the 'Observation' returned by the tool. It essentially creates a step-by-step audit trail of the agent's logic.

Question 5

What is the Plan-and-Execute agent pattern?

Accepted Answer

Plan-and-Execute is a multi-step agent architecture that splits the workload: a 'Planner' agent creates a step-by-step master plan, and an 'Executor' agent completes each step sequentially. See: Plan-and-Execute Agent. While ReAct decides its next action dynamically step-by-step, it can get stuck in loops on very complex tasks. Plan-and-Execute mimics human project management. First, it generates an entire DAG (Directed Acyclic Graph) of sub-tasks. Then, it executes them one by one, occasionally calling a 'Replanner' to adjust the remaining steps based on new information.

Question 6

What is tool use (function calling) in LLMs, and how does it enable agents?

Accepted Answer

Tool use (Function Calling) is a native API feature where an LLM is provided with a JSON schema of available functions and, instead of generating text, outputs a structured JSON object containing the function name and arguments to execute. See video: AI Engineering Explained. Before native function calling, developers had to parse raw text using Regex to extract commands. Now, models are fine-tuned to recognize when a tool is needed. The application sends a list of available tools to the API. If the model decides it needs weather data, it pauses its text generation and outputs a structured request: `{'name': 'get_weather', 'arguments': {'location': 'NYC'}}`. The application runs the code and feeds the result back.

Question 7

What is the difference between single-agent and multi-agent systems?

Accepted Answer

A single-agent system relies on one massive prompt managing all tools and reasoning. A multi-agent system divides a complex task among multiple specialized agents, each with its own distinct persona, tools, and system prompt, working collaboratively. See: Multi-Agent Systems. In a single-agent setup, giving the model 50 tools and asking it to be a researcher, coder, and QA tester simultaneously will overwhelm its attention mechanism. Multi-agent systems (like CrewAI or AutoGen) solve this by creating a 'Researcher Agent' with search tools, a 'Coder Agent' with execution tools, and a 'Manager Agent' to orchestrate them. They pass messages to each other like a human team.

Question 8

What is Model Context Protocol (MCP), and how does it standardize tool integration?

Accepted Answer

The Model Context Protocol (MCP) is an emerging open standard that provides a unified way for AI models to securely connect to external data sources and tools without requiring custom integration code for every new application. See video: AI Engineering Explained. Historically, integrating a new tool (like GitHub or Slack) into an agent required writing custom API wrappers. MCP acts like a 'USB-C cable for AI'. It standardizes the interface. An application exposes its capabilities via an MCP server, and any MCP-compliant AI agent (like Claude Desktop) can instantly connect to it, discover its tools, and read its context securely.

Question 9

How do you handle agent failures and implement error recovery?

Accepted Answer

Agent error recovery is implemented by catching tool execution exceptions in the application harness, formatting the error stack trace into a readable string, and feeding it back into the agent's prompt with an instruction to fix the parameters. Agents will inevitably hallucinate invalid arguments or use incorrect syntax. Robust systems do not crash when this happens. Instead, the application catches the 400 Bad Request or Python exception, creates a system message (`Observation: Tool failed with error: [Error Details]. Please review your arguments and try again.`), and loops back to the agent.

Question 10

What is an agent loop, and how does it decide when to stop?

Accepted Answer

An agent loop is a continuous execution cycle (e.g., while true: Think -> Act -> Observe). It stops either when the LLM generates a specific 'Finish' action, or when the application hits a predefined limit (max iterations or max time). See: AI Agent Explained. The core of any agent framework (like LangGraph) is a state machine looping through execution nodes. The LLM decides to stop by selecting a dedicated tool (e.g., `FinalAnswerTool`) indicating the goal is met. The host application acts as a circuit breaker, monitoring the loop and forcefully terminating it if the agent exceeds budget or gets stuck.

Question 11

What is Context Engineering in agentic systems?

Accepted Answer

Context Engineering is the practice of dynamically constructing, filtering, and injecting exactly the right information into the LLM's context window at runtime, ensuring the agent has the exact data it needs without being overwhelmed by noise. See: Context Engineering. Unlike Prompt Engineering (which focuses on instructions), Context Engineering focuses on data payload management. Because context windows are limited and expensive, engineers must implement dynamic truncation, vector retrieval, summarization, and variable injection to ensure the agent receives a pristine, high-signal data payload for its current task.

Question 12

How do you evaluate and test AI agents?

Accepted Answer

Evaluating AI agents requires a multi-faceted approach: assessing the final output quality, the efficiency of the trajectory (number of tool calls used), the accuracy of tool selection, and the agent's ability to recover from simulated errors. Testing an agent is fundamentally different from testing an LLM. An LLM is evaluated on text quality. An agent is evaluated on behavior. Frameworks use 'Trajectory Evaluation' to analyze the steps the agent took. If Agent A reaches the goal in 3 tool calls, and Agent B reaches it in 15 tool calls, Agent A is superior. Testing involves creating deterministic sandbox environments where tool outputs are mocked.

Question 13

What are the security risks of agentic systems, and how do you mitigate them?

Accepted Answer

The primary risks are Prompt Injection hijacking the agent's goal, unauthorized access via over-permissioned tools, and destructive autonomous actions (e.g., deleting a database). Because agents have agency (the ability to act), their blast radius is massive. Mitigation requires implementing Human-in-the-Loop (HITL) approval gates for destructive actions, running the agent in ephemeral Docker sandboxes, enforcing the Principle of Least Privilege on API keys, and scanning inputs for injection attempts.

Question 14

How do you design and define tools for an AI agent?

Accepted Answer

Tools should be designed with highly descriptive names, strictly typed JSON schemas with detailed parameter descriptions, and robust error handling that returns readable error strings rather than crashing. See video: AI Engineering Explained. An LLM relies entirely on the docstrings and parameter descriptions to understand a tool. A tool named `fetch_data()` is terrible. A tool named `get_customer_order_by_id(order_id: string)` with a description of 'Fetches the shipping status of a customer order' is excellent. The tool must return clean text or JSON, not massive HTML blobs that blow up the context window.

Agentic Systems Interview Prep Portal