Skip to content

🧠 Lesson 0-3: Thinking Like an Agent

Learning Objectives

By the end of this lesson, you will be able to:

  • Decompose complex goals into agent-compatible subtasks using agentic reasoning frameworks
  • Apply behavior trees and chain-of-thought methods to structure multi-step processes
  • Recognize common failure points and cost issues in autonomous workflows
  • Sketch reflective loops and feedback cycles to increase agent reliability and robustness

🎯 1. Introduction: Why Agentic Thinking Matters

Successful agents don't just execute pre-defined scripts; they plan ahead, adapt, and reflect at every step. This mindset goes beyond classic procedural code, asking:

  • What is the end goal?
  • What are the minimal, modular steps needed?
  • What tools or reasoning methods apply for each step?
  • How and when should the agent check its own progress or results?

Key Principle

Agents break large goals into "bite-sized" subtasks, select the right tool for each, act, observe, and adjust based on results.

🔍 2. Problem Decomposition: Goal → Subtasks

To act autonomously, agents must transform broad goals into a sequence (or tree) of subtasks.

Example Goal Decomposition

Goal: "Book a meeting with Alice and Bob next week."

Decomposition: - Search both users' calendars for availability - Propose potential times - Draft invite and confirmation message - Send invite and await acceptance

Diagram: 

[image: goal-subtask-tree]
A tree with root "Book meeting" branching to "Search calendars", "Propose times", "Draft invite", "Send/await acceptance"

Did you know?

The process of breaking goals into ordered subtasks is called "task decomposition" and is a core building block of planning agents.

🧩 3. Reasoning Frameworks: Behavior Trees & Chain-of-Thought

3.1 Behavior Trees

Used widely in robotics and games, behavior trees structure agent logic into nodes:

  • Selector Nodes: Choose which branch to attempt, based on conditions
  • Sequence Nodes: Execute children in order; fail if any step fails
  • Action Leaves: Invoke a tool or LLM call

Mini Behavior Tree

Book Meeting (Selector)
 ├─ Check calendars (Action)
 ├─ Find slots (Action)
 ├─ Propose time (Action)
 ├─ Draft invite (Action)
 └─ Send invite (Action)

3.2 Chain-of-Thought Prompts

Modern LLM-powered agents benefit from thinking aloud before acting:

  • "First, I'll check Alice's schedule. Next, I'll check Bob's. Then, I'll look for overlapping free slots. Finally, I'll send an invite."
  • This step-wise reasoning helps the agent avoid skipping or muddling steps, providing transparency and traceability.

Chain-of-Thought Benefits

Chain-of-thought (CoT) reduces hallucinations and improves reliability in complex, multi-step tasks.

🔄 4. Reflective Loops and Self-Correction

Agents inevitably encounter uncertainty—missing info, unexpected outputs, or tool errors. Reflection is the art of checking results, diagnosing issues, and deciding what to do next.

Signs your agent needs reflection:

  • Stuck in infinite loop (keeps retrying same step)
  • Repeats failed actions ("Cannot find calendar. Cannot find calendar…")
  • Outputs unclear or partial results

Mitigation Strategies:

  • Observation logging: Write each action/observation to memory for inspection
  • Conditional checks: If output is blank, error, or duplicate, trigger different plan or escalate
  • Feedback cycle: Use results of one step to alter subsequent steps ("if time slot not found, search next week")

🚀 5. Worked Example: Agent for Data Pipeline

Data Pipeline Agent

Goal: "Ingest a CSV file, clean the data, compute summary statistics, and email results."

Agentic Decomposition:

Subtasks: - Load CSV from given path - Clean data (remove nulls/outliers) - Compute summary statistics (mean, median) - Generate summary report - Draft and send email with results attached

Behavior Tree: 

Ingest & Report Pipeline (Sequence)
 ├─ Load CSV (Action)
 ├─ Clean Data (Action)
 ├─ Compute Stats (Action)
 ├─ Generate Report (Action)
 └─ Send Email (Action)

Chain-of-Thought Example: - "I will start by loading the CSV file… The file loaded. I see there are missing values. I'll remove rows with missing values… Now compute stats—mean, median, std. Compose a brief summary. Attach to an email and send to the recipient."

Case Study

In production, such an agent failed when the CSV was missing columns; self-check logic added: "If expected columns not found, send error notification."

💻 6. Mini-Project Prompt

Onboarding Agent Challenge

Task: Design a high-level agent flow for "Onboard a new employee" (inputs: name, start date, department).

Step 1: Sketch possible subtasks (e.g., create IT account, send onboarding document, schedule intro meeting, send welcome email).

Step 2: Draw a behavior tree or write a chain-of-thought outline explaining the agent's reasoning.

Optional: Write a Python pseudocode list representing these subtasks and how the agent would iterate through them.

❓ 7. Self-Check Questions

Knowledge Check

  1. Why should agents "think aloud" using chain-of-thought before acting?
  2. Name one benefit and one risk of automatically decomposing high-level goals.
  3. How does a behavior tree help prevent infinite loops or critical step omissions?
  4. What's the first thing your onboarding agent should check before proceeding?

Phase Complete!

Phase 1: Tools, Prompts & Memory →

Ready to start building? Learn Python integrations, prompt engineering, and memory systems.