Concept Review - Agents
Conceptual Questions
Q: What is an AI Agent? An AI Agent is a system that uses a Large Language Model (LLM) as its "brain" to perceive its environment, reason through complex goals, and execute actions autonomously using external tools to achieve a specific outcome. Unlike a chatbot (single input → single output), an agent pursues a goal over multiple steps - it acts, receives feedback, and adapts rather than just generating one response.
Q: What are the key features of AI Agents?
- Autonomy: Act independently without constant human intervention
- Goal-Oriented: Driven by objectives, aiming to maximize success based on defined metrics
- Perception & Observation: Gather information from the environment to understand context
- Reasoning & Rationality: Analyze data, identify patterns, and make informed decisions
- Planning: Develop strategic plans to achieve goals
- Acting & Proactivity: Take initiative and execute actions based on decisions, anticipating events rather than just reacting
- Adaptability: Adjust strategies and learn in response to new circumstances
- Collaboration: Work effectively with humans or other AI agents through communication and coordination
Q: What is the agent loop (Perceive → Reason → Act)? Perceive: receive input - user message, tool result, or environmental state. Reason: use the LLM to decide what to do next (which tool to call, or whether to respond). Act: execute the chosen action - call a tool, run code, search the web, etc. The loop repeats until a stopping condition is met (task complete, max steps reached, confidence threshold).
Q: What is tool use / function calling? Function calling allows the LLM to request the execution of a predefined function by outputting structured JSON (function name + arguments) instead of text. The application executes the function and returns the result to the LLM. This turns the LLM from a text generator into something that can interact with the world: search APIs, databases, code interpreters, external services.
Q: What is the difference between tool use and MCP? Tool use (function calling) is a mechanism - the LLM outputs a function signature, your application calls it. Each integration is custom. MCP is a protocol that standardizes how tools and data sources are exposed to LLMs - a universal interface so any MCP server can work with any MCP-compatible host. MCP is the scalable, ecosystem-level answer to the N×M integration problem.
Framework-Specific Questions
Q: What is the ReAct pattern? Reasoning + Acting interleaved. The agent generates a Thought (internal reasoning about what to do), then an Action (tool call), then receives an Observation (tool result), then reasons again. This cycle continues until a Final Answer is reached. Compared to pure CoT, ReAct has access to real-world information via tools, making it far more capable for information-retrieval and multi-step tasks.
Q: When would you choose LangGraph over LangChain? Use LangGraph when: (1) your agent needs cycles - revisiting earlier steps, retrying, looping; (2) you need persistent, checkpointed state across a long conversation; (3) you want explicit control over agent flow with branching logic; (4) you're building multi-agent systems with defined coordination. LangChain is sufficient for linear pipelines. LangGraph is built on top of LangChain - it's an extension, not a replacement.
Q: What is CrewAI's role-based model and when is it useful? In CrewAI, each agent has a defined Role, Goal, and Backstory - this primes the LLM to behave like a specific persona (researcher, writer, critic). Agents are assigned Tasks and can delegate to each other. This model shines when a problem naturally decomposes into distinct professional roles. It's more intuitive to design than explicit graph-based flows but offers less low-level control.
Architecture and Design Questions
Q: How do agents maintain state and memory? Four types: (1) In-context - information in the current prompt window (short-lived). (2) External storage - key-value stores, databases, vector stores (persistent). (3) Episodic memory - past conversation summaries retrieved at the start of new sessions. (4) Semantic memory - a knowledge base the agent can query. Most production systems combine all four.
Q: What are the main failure modes of LLM agents? (1) Tool call loops - agent keeps calling the same tool without progress. (2) Hallucinated tool arguments - LLM generates invalid parameters. (3) Context overflow - accumulated observations exceed the context window. (4) Overconfident stopping - agent decides it's done before the task is complete. (5) Cascading errors - early tool call failure derails the entire plan. Mitigation: max-step limits, result validation, structured error handling, human-in-the-loop checkpoints.
Q: What is a multi-agent system and when do you need one? Multiple specialized agents collaborating on a task - an orchestrator decomposes the task and delegates to subagents (researcher, coder, reviewer). Needed when: (1) parallelism - subtasks can run concurrently; (2) specialization - different parts require different personas or tools; (3) context management - each subagent handles its own context window. The cost is coordination overhead and harder debugging.