Level 1
0 XP
🌟 Start Here

Welcome to Agentic Engineering

Learn to design, direct, and build AI agent systems — no coding experience required.

You've probably used AI chatbots like ChatGPT or Claude. You type a question, it answers. That's useful — but it's just the beginning.

Agentic Engineering is the next level. Instead of just chatting with AI, you learn to design systems where AI agents work autonomously — using tools, following plans, checking their own work, and even collaborating with other agents.

💡
Key insight: The difference between a chatbot and an agent is like the difference between asking someone a question and hiring them to complete a project.

🎬 Optional: Build Your First AI Agent (No Coding)

This 25-minute video walks you through building a real AI agent from scratch without writing any code. Watch it now for motivation, or come back after completing the course.

25:58 · Futurepedia · 3.6M views · Optional — watch for inspiration or revisit after the course.

Chatbot vs. Agent

💬 Chatbot
  • You ask, it answers
  • One turn at a time
  • No tools, no actions
  • Forgets everything next session
  • You do all the thinking
🤖 Agent
  • You describe a goal, it works toward it
  • Multiple steps, plans ahead
  • Uses tools (search, edit, run code)
  • Can save and reuse context when configured
  • It thinks, plans, and executes

The 5 Levels of Mastery

This course follows a progression from beginner to expert:

🗣

L1: Prompter

Communicate clearly with AI agents

L2: Automator

Build repeatable workflows

🛠

L3: Architect

Design context, tools, and specs

🔧

L4: Engineer

Build reliable agent systems

🏆

L5: Orchestrator

Coordinate multi-agent teams

Quick Check: What makes something an "agent"?

Level 1: Foundations

Set Your Agent Up to Succeed

The #1 insight: it's not about having a smarter AI — it's about giving it the right environment to work in.

The Harness Analogy

Imagine you have the smartest student in the world taking an exam. But you:

  • Give them no scratch paper
  • Don't let them ask clarifying questions
  • Never check their work before submitting
  • Give them all questions at once with no breaks

They'd perform terribly — not because they're dumb, but because the environment set them up to fail.

📈
Real data: Researchers found that improving the "harness" (the system around an AI) can improve performance by 6x — without changing the AI model at all. Same brain, better environment = dramatically better results.

What Is a "Harness"?

A harness is everything around the AI that helps it succeed:

📝

Instructions

Clear prompts that tell the AI exactly what to do and how

🔧

Tools

Actions the AI can take (search files, run commands, edit code)

🚫

Guardrails

Rules that prevent mistakes (don't delete files, ask before pushing)

🔍

Feedback

Ways for the AI to check its own work (tests, screenshots, reviews)

Scenario: Fix the Harness

🎯 Challenge

A student asks an AI agent: "Make my website look better." The agent changes random colors, breaks the layout, and doesn't check the result.

What harness improvement would help most?

Researchers at Stanford and MIT took this idea further: what if the agent could improve its own harness?

They built a system where an AI would read its previous performance, analyze what went wrong, then rewrite its own instructions and tool configurations. The result? A small, cheap model (Haiku) with a self-optimizing harness ranked #1 in its weight class — beating models 10x its size.

This is called a meta-harness: a harness that improves itself.

Level 2: Automator

Your First Agent Workflow

Move beyond one-shot prompts. Learn to chain steps into powerful automated workflows.

The Explore → Plan → Execute Loop

Great agent work follows a pattern:

🔎
Explore
📝
Plan
Confirm
💻
Execute
🔍
Verify

Each step builds on the previous one — explore first, then plan, confirm, execute, and verify.

Three Workflow Patterns

🤝

Guided Mode

Agent explores and plans, but asks YOU before making changes. Safest for beginners.

🔄

Iterative Mode

Agent makes changes, checks its own work (tests/screenshots), and fixes issues automatically.

🏃

Autonomous Mode

Agent runs independently for extended periods. Most powerful but needs good guardrails.

Real Examples

# Guided: Agent asks before acting
You: Add a dark mode toggle. Explore the CSS first, make a plan, and ask me before changing anything.
# Iterative: Agent checks its own work
You: Build this UI from the mockup. Take a screenshot after each change to compare with the design. Iterate until it matches.
# Autonomous: Agent runs independently
You: Run the test suite. For each failing test, fix the code and re-run until all tests pass.

Which workflow pattern should you use when you're not sure what the agent will do?

Level 2: Automator

Context is King

The more context you give an agent, the smarter it becomes. Learn to engineer context like a pro.

The Context Window: An Agent's Working Memory

AI agents have a context window — think of it like a desk. Everything the agent needs to "see" must fit on this desk. If the desk is cluttered with irrelevant stuff, the agent gets confused. If the desk is organized with exactly what's needed, the agent works brilliantly.

📚
The exam analogy: On a 3-hour exam, don't try to write one 3-hour essay. Do one problem, check your work, take a break, then start the next. Your focus stays sharp. Same with agents — break big tasks into small, focused chunks.

CLAUDE.md: The Briefing Document

A CLAUDE.md file is like giving a new employee an orientation packet. It tells the agent:

  • What this project is about
  • What tools and commands to use
  • What rules to follow
  • What style to write in
# My Science Fair Project
## Rules
- Write at a high school reading level
- Always cite sources in MLA format
- Ask before changing my bibliography
## My Files
- essay/ contains my paper draft
- notes/ contains my research notes
- sources/ contains articles I've found

The Context Hierarchy

🏢 Enterprise / School Policy Rules everyone must follow (e.g., "never share student data")
👤 Personal Preferences Your individual settings (e.g., "call me Alex, be concise")
📁 Project Context Shared with your team (e.g., "use Python 3.11, run pytest")
📄 Task-Specific Context Just for this task (e.g., "this file handles login")

You want an agent to help with your biology project. Which context would help it most?

Level 3: Architect

Think Before You Build

The spec-first approach: planning prevents disasters. Don't "vibe code" — write specs.

Vibe Coding vs. Spec-Driven

💥 Vibe Coding

"Hey AI, build me a social media app"

  • No plan, just hope
  • AI guesses what you want
  • Hours of rework
  • Result rarely matches vision
📝 Spec-Driven

"Let's write a spec together first"

  • Clear requirements up front
  • AI and you agree on the plan
  • Catch problems early
  • Result matches your vision
🏗
Contractor analogy: Don't hire a builder and say "make it nice." Draw a blueprint first, agree on every detail, get sign-off — THEN build. Saves everyone time and money.

The 4-Step Spec Process

🤔
1. Understand
📝
2. Spec
💻
3. Implement
😈
4. Verify

Example Spec Template

## Goal
Build a quiz app for my biology class
## Requirements
- 10 multiple choice questions about cells
- Show score at the end
- Works on mobile phones
- No backend needed (all client-side)
## Out of Scope
- No user accounts
- No leaderboard (maybe later)
## Success Criteria
- All questions display correctly on mobile
- Score calculates accurately
- Loads in under 2 seconds

Why is "Out of Scope" important in a spec?

Level 3: Architect

Designing Agent Tools

The tools you give an agent determine what it can do. Learn to pick and design the right ones.

Tools = Superpowers

Without tools, an AI can only talk. With tools, it can act. Each tool gives the agent a new capability:

🔍

Search

Find files, search the web, look up information

Edit

Create and modify files, write documents

Execute

Run programs, scripts, commands

📸

Observe

Take screenshots, read outputs, check results

💬

Communicate

Ask users questions, send messages, create PRs

🧠

Remember

Save notes, update memory, track progress

The Tool Design Principle

🎯
Anthropic's rule: Design tools that match what the AI is naturally good at — not what humans assume it needs. Watch how the AI uses tools, then adjust.
🚫 Bad Tool Design

A tool with 50 required parameters that the AI gets wrong half the time

✓ Good Tool Design

A simple tool with smart defaults that the AI uses correctly 95% of the time

You're building an agent to help write essays. Which tool would help it most?

Level 4: Engineer

Agent Memory & Learning

How agents remember things across sessions and get smarter over time.

The Memory Problem

Many AI agents only know what's in the current conversation. Once the session ends, they can lose track of everything you discussed. It's like having a brilliant tutor who doesn't take notes — every session, you start explaining from scratch.

Memory systems solve this by saving important information between sessions.

Types of Agent Memory

📝

Flat Memory

Simple notes saved to a file. Like sticky notes. Easy but messy at scale.

🔗

Graph Memory

Connected web of facts. "Task B depends on Task A." The agent sees relationships.

📚

Episodic Memory

Records of past sessions. "Last time I tried X, it failed because Y."

Memory in Practice

# Tell the agent to remember something
You: # Remember: this project uses MLA citation format, not APA
# Check what the agent remembers
You: /memory
Loaded memories:
- Project uses MLA citations
- Preferred language: concise
- Always run tests before committing

What's the advantage of graph memory over flat memory?

Level 4: Engineer

Agent Safety & Guardrails

With great power comes great responsibility. Learn to keep agents safe and controlled.

Why Safety Matters

An AI agent with tools can take real actions — edit files, run programs, send messages, even delete things. Without guardrails, a small mistake can cascade into big problems.

🛑
Real example: An agent told to "clean up the project" once deleted important files it thought were unnecessary. The fix? Add a guardrail: "never delete files without asking first."

Types of Guardrails

🔒

Permission Levels

Some actions need your approval (delete, push, install). Others are auto-approved (read, search).

🚫

Blocklists

Actions that are NEVER allowed, no matter what (e.g., "never access /private/").

📦

Sandboxes

Isolated environments where the agent can experiment without affecting real systems.

💰

Cost Limits

Spending caps that stop the agent from burning through too many resources.

The Permission Pyramid

Always Allowed
Read files, search, list directories
Ask First
Edit files, run scripts, install packages
Careful
Delete, push to git, send messages
Never
Access secrets, bypass safety

An agent wants to delete 50 files it thinks are "unused." What should happen?

Level 5: Orchestrator

Multi-Agent Teams

One agent is good. Multiple agents working together is transformational.

The Software Company Analogy

Real software companies don't have one person do everything. They have specialists:

📋
PM
🛠
Architect
💻
Engineer
🔎
QA
🚀
Deploy

Multi-agent systems replicate this: different AI agents take on different roles, passing work between each other like a relay team.

Why Multiple Agents?

🎯

Specialization

Each agent focuses on one job and does it well

Parallelism

Multiple agents can work on different tasks simultaneously

🔍

Review

One agent can check another's work (like peer review)

Real Example: MetaGPT

MetaGPT simulates an entire software company with AI agents. Give it one prompt like "build a snake game" and:

  1. Product Manager Agent writes requirements
  2. Architect Agent designs the system
  3. Engineer Agent writes the code
  4. QA Agent tests for bugs

Each agent has clear inputs and outputs. The PM's document becomes the Architect's input. The Architect's design becomes the Engineer's specification. No confusion about responsibilities.

🏆
Result: MetaGPT (67,000 GitHub stars) produces better code than a single agent because specialization + review catches errors that one agent alone would miss.

What's the main risk of multi-agent systems?

Level 5: Orchestrator

Design Your Own Agent System

Put everything together. Design an agent system from scratch using the Agent Architecture Canvas.

The Agent Architecture Canvas

Every well-designed agent system answers these 8 questions. Fill in each section to design yours:

1. Goal

What problem does your agent solve? Who benefits?

2. Tools

What actions does your agent need to take?

3. Context

What information does it need to work well?

4. Workflow

What pattern does it follow? (Guided/Iterative/Autonomous)

5. Memory

What does it need to remember across sessions?

6. Safety

What guardrails prevent problems?

7. Verification

How do you know it's working correctly?

8. Human-in-the-Loop

When should a human step in?