⬡ Overview
↻ The Loop
◈ Agent Types
⌖ Support Agent
⚑ Failure Modes
◇ Knowledge Check
★ Results
0 / 4
Hands-On Lab  ·  Module 3  ·  BAA 3.0

How Agents
Work

From classical AI theory to modern LLM stacks — trace the perceive–reason–act loop from first principles, classify the four agent architectures, and diagnose the failure modes that break real systems.

Learning Objectives
01
Identify the classical Perceive–Act loop and explain why iteration is what creates autonomy.
02
Compare the four classical agent architectures: Reactive, Goal-Based, Utility-Based, and Learning.
03
Map Perception, Reasoning, Tools, Action, and Memory onto a real customer service agent scenario.
04
Diagnose the four structural failure modes and name a prevention strategy for each.
Your Lab Roadmap
1
The Classical Loop
Perceive → Reason → Act → Observe → Repeat — click each step to explore
2
Four Classical Agent Types
Reactive · Goal-Based · Utility-Based · Learning — comparison table
3
Support Agent Anatomy
Drag 15 behaviors into the correct loop stage for a CS agent
4
Structural Failure Modes
Diagnose 12 symptoms across 4 failure categories
5
Knowledge Check
7 questions · agent loop mastery · 70% to pass
—
Your Results
Score breakdown · key takeaways
Section 01

The Classical Agent Loop

Every agent — classical or modern — runs this loop. Click each step to understand what happens at that moment.

💡 Key insight: The loop is what separates agents from generators. Without iteration, there is no autonomy — only prediction.
Step 01
👁️
Perceive
Step 02
🧠
Reason
Step 03
⚡
Act
Step 04
📊
Observe
Step 05
🔄
Repeat
↩  REPEAT UNTIL GOAL IS MET  ↩
PERCEIVE — The agent reads its environment. For a modern LLM agent, this means receiving a user message, an API event, a scheduled trigger, or a tool result. Without perception there is nothing to act on — this is the entry point of every loop iteration.
Why the Loop Matters
🔁

Correct Mistakes

If the first action fails, the agent tries a different approach instead of stopping.

🌊

Adapt to Change

Environment shifts require re-evaluation and new plans — only possible through iteration.

🎯

Achieve Autonomy

Multi-step goal pursuit requires continuous operation — not one-shot text generation.

📈

Learn from Results

Feedback from actions informs better decisions. The loop transforms prediction into behavior.

Section 02

Four Classical Agent Types

From classical AI theory (Russell & Norvig). Click each card to explore the architecture. Modern LLM agents are hybrid systems combining elements from all four.

01
⚡

Reactive Agents

Stimulus → Response. No memory, no planning.

Tap to expand ↓

No internal model of the world. Immediate, deterministic response to input. Example: A thermostat — temperature drops below threshold → turn on heater. No memory, no learning, no planning. Fast and reliable but brittle — can't handle anything outside its trigger conditions.
02
🎯

Goal-Based Agents

Compare state vs. goal. Introduce planning.

Tap to expand ↓

Evaluate actions based on: "Does this move me closer to the goal?" Goal-based agents introduce planning — the ability to evaluate future states and select actions that advance toward a defined objective. First type capable of genuine multi-step reasoning.
03
⚖️

Utility-Based Agents

Optimize trade-offs. Maximize a utility function.

Tap to expand ↓

Instead of binary goals, utility-based agents optimize trade-offs. Example: minimize cost ⚖️ maximize speed. They find the optimal balance — not just "yes/no" on a single goal. This is the foundation of modern recommendation and routing systems.
04
🧬

Learning Agents

Update model from experience. Adapt over time.

Tap to expand ↓

Four components: Performance Element (selects actions), Learning Element (improves from feedback), Critic (evaluates outcomes), Knowledge Base (stores experience). Most advanced classical type — closest to modern LLM agents.
Comparative Table

Review before the quiz — you'll be asked which capabilities each type has.

Agent TypeMemoryPlanningOptimizationLearning
Reactive✗✗✗✗
Goal-Based~✓~✗
Utility-Based✓✓✓✗
Learning✓✓✓✓
Modern LLM Agent✓✓✓✓
🤖 Modern LLM agents are hybrid systems. The architecture wrapping the LLM determines which capabilities are active — the model alone decides nothing.
Section 03 · Lab Activity

Support Agent Anatomy

Diagram the agent loop for a customer service responder. Drag each behavior into the correct stage — Perception, Reasoning, Tools, Action, or Memory.

Scenario: A customer submits a ticket: "My order hasn't arrived — it's been 10 days and I need it urgently." A fully autonomous support agent handles the entire resolution loop without human intervention. Your job is to map what happens at each stage of the loop.
Unassigned behaviors — drag to the correct stage
On mobile? Tap to assign
Section 04

Structural Failure Modes

Most agent failures are architectural — not model errors. Drag each symptom into the failure mode it belongs to. Tap a header to reveal the prevention strategy.

⚠️ 4 Failure Modes: Poor Task Decomposition · Tool Misalignment · Feedback Ignorance · Over-Autonomy
Undiagnosed symptoms — drag to assign
On mobile? Tap to assign
Section 05

Knowledge Check

Test your understanding. You need 70% or higher to mark this section complete.

Question 1 of 7Score: 0 / 0
Your Results

Lab Complete