LLM Fundamentals | Visual Explainer

Agents can be classified by how they decide and act: from simple rule-based reflexes to multi-agent systems. Choosing a type depends on whether the task is open-ended, multi-step, or should improve over time.

Types of agents (travel-booking examples)

Type	Description	Example
Simple Reflex	Immediate actions from predefined rules.	Travel agent forwards complaints to customer service by rule.
Model-Based Reflex	Keeps an internal picture of the world (e.g. past prices, inventory) and acts using that picture, not just the current snapshot.	Uses historical flight prices to highlight routes that dropped a lot in price or spiked — so you see 'good deal' or 'book soon' based on how the world (prices) changed.
Goal-Based	Creates plans to reach a specific goal.	Books a journey by deciding flights, car, transit from A to B.
Utility-Based	Weighs tradeoffs (e.g. cost vs convenience) numerically.	Maximizes utility when choosing between cheap vs convenient flights.
Learning	Improves over time from feedback.	Uses post-trip surveys to adjust future recommendations.
Hierarchical	Higher-level agent breaks work into subtasks for lower-level agents.	Cancel trip → cancel flight, hotel, car; each subtask handled by a sub-agent.
Multi-Agent (MAS)	Multiple agents work together or compete.	Cooperative: one books hotel, one flights. Competitive: agents share a calendar.

What does “model of the world” mean?

A simple reflex agent only looks at the current input (e.g. “user asked for flights”) and applies a rule. A model-based reflex agent keeps an internal picture of how things change over time (e.g. historical prices, yesterday’s inventory). It uses that picture to decide — for example, “this route dropped a lot in price, so surface it as a deal” or “this one is spiking, warn the user.” The “model” is that internal representation of the world, not just the latest snapshot.

When AI agents shine

Open-ended problems

Steps can't be hardcoded; the LLM decides what to do next.

Multi-step processes

Task needs tools or information over many turns, not a single shot.

Improvement over time

Agent gets feedback from environment or users and gets better.

Example: Travel agent by type

A simple reflex agent forwards complaints to customer service by rule. A model-based reflex uses past data to highlight price drops or spikes. A goal-based agent plans flights + hotel + car to reach a destination. A learning agent uses post-trip feedback to improve future recommendations.

📋 Chapter 2: Types of Agents & When to Use Them