AI Engineering Visualized: Agents, The Course

One harness, built module by module.

I wrote this course for developers who have shipped a language-model feature and now have to let it act: issue a refund, cancel an order, change an address, wait for a courier, hand a case to a person. A stronger model won't stop it refunding someone else's order. The code around the model will, and that's what we draw and build here, one part at a time.

What you'll build

You start with an assistant that can be talked into refunding someone else's order. By the end, your agent:

• runs in a harness where the model proposes and code authorizes, with the customer's identity carried from the session, never from the model
• calls tools with typed specs, and caps and previews on anything that moves money
• checks its plans in code and repairs only the branch that broke
• passes verifiers it can't talk its way past before anything is committed
• moves through a state machine with checkpoints, idempotency keys, a ledger and named stop reasons
• waits for couriers, asks people for approval and hands cases off cleanly
• keeps memory you can correct, a context budget and retrieval it chooses itself
• connects to MCP servers it can trust, and survives loops, retry storms and failing dependencies
• is graded on outcomes and trajectories in a sandbox, across repeat runs, before a release gate lets it ship

What's inside

• 17 modules, plus a short Start Here module that sets up the code
• 17 video lessons, 6 to 7 minutes each, so nearly 2 hours of video in all
• 100+ hand-drawn doodles, each one a named mental model like the Agent Harness or the Clean Handoff
• 17 one-page doodle cheat sheets
• 17 coding exercises, each with a prediction question, the expected output and a worked solution
• 17 quizzes. Conceptual questions only. Retake as often as you like; your best result counts.
• A certificate of completion: pass with 70% or more, or 90% or more for a distinction

How each module works

Watch the video. Read the lesson. Keep the cheat sheet. Do the exercise: predict what the code will print, run it, compare. Take the quiz. That's it, at your own pace, with no start date and no deadline.

The code

Everything runs on Python 3.10 or newer, with one requirements file to install. A built-in stub model means you don't need an API key or a paid account, and every exercise has an expected output you can check against. If you'd like to try a real model, you can point the same code at Ollama on your own machine.

Who it's for

Developers who can read Python and have taken Volume 1, or covered the same ground: message lists, tools, evals and a basic agent loop. You don't need a machine learning background.

Who it's not for

If you want a tour of agent frameworks or want to train models, this isn't that course. It builds a plain loop so that every check is code you can read.

Modules

0 Start Here · 1 The Agent Harness · 2 Tool Engineering · 3 Planning · 4 Verification Before Commit · 5 State Machines, Checkpoints and Stop Reasons · 6 Side Effects · 7 Long-Running Tickets · 8 Human Oversight · 9 Agent Memory · 10 Context Engineering · 11 Agentic RAG · 12 MCP · 13 Tool Trust · 14 Multi-Agent Systems · 15 Agent Failure Engineering · 16 Agent Evaluation · 17 Agent Reliability

The course and the book

The course follows the book AI Engineering Visualized, Volume 2: Agents (Vol 2), chapter for chapter. If you already have the book, the course adds the videos in each module, the one-page doodle cheat sheets, the prediction questions, the quizzes and the certificate, with your progress tracked as you go.