What Is an LLM Agent?

When a user asks an LLM-powered assistant to research a product, compare prices across three different retailers, and then place an order for the cheapest option, a single prompt-in, completion-out call cannot get the job done. The task demands multiple steps: fetching live pricing data from external sources, making conditional decisions based on what comes back, and executing a purchase action. A standard LLM call, no matter how capable the model, produces one static response and stops. It has no ability to go out into the world, gather information, evaluate it, and take further action based on what it finds.

This is the gap that LLM agents fill. An LLM agent is a system where the language model itself decides what action to take next, observes the result of that action, and continues iterating until the goal is met. Instead of a developer scripting every step in advance, the model reasons its way through the problem dynamically. Real-world agentic workflows already operate this way: autonomous customer-support bots that look up order status, issue refunds, and escalate edge cases without human intervention; code-generation pipelines that write code, run tests, read error messages, and self-debug; and research assistants that refine search queries across multiple iterations until they surface the most relevant results.

This lesson builds the foundational mental model you need before working with agents hands-on. You will first see how agents differ from deterministic chains, then explore the perceive-reason-act loop that drives all agent behavior, and finally understand why this loop is the prerequisite for the tools and memory components covered in the next lesson.

Deterministic chains vs. LLM agents

To ...