Core Components of an LLM-Based AI Agent and How They Interact

What are the components of an AI agent? Name each component, explain what it does, and show how the components work together when the agent carries out a multi-step task. ```hint Follow one task Pick a concrete multi-step task and trace what the agent needs at each step. Each need points to a component. ``` ```hint What makes it stop Think about what ends the agent's loop, and what prevents it from taking a harmful or expensive action along the way. ``` ### Clarifying Questions - Does "AI agent" here mean a system built around a large language model, or an agent in the broader reinforcement-learning sense? - Should the agent act autonomously, or with a person approving some actions? - One agent, or several agents that cooperate? ### What a Strong Answer Covers - The model as the reasoning core, and the instructions that define the agent's goal and behavior - Tools with typed interfaces, and how the model invokes them - Short-term and long-term memory, and how context is managed - Planning and the control loop, including stopping conditions and budgets - Guardrails, permissions, evaluation and observability - A concrete walk-through showing how the components interact, and the link to the classical perceive-decide-act view ### Follow-up Questions - How would you keep an agent from looping forever or exhausting its budget? - A tool returns text that contains instructions addressed to the agent. What is the risk, and how do you defend against it? - How would you evaluate an agent whose tasks have many valid action sequences? - When would you build a fixed workflow instead of an autonomous agent?