The Handbook
Eleven illustrated sections on who decides the next step at run time — writing code, prompts, context, loops, graphs, harnesses, evaluation and security. Every model interaction here is a scripted simulation.
Read the handbook→Every program is a list of steps. The only question is who picks them.
A complete course on building systems that use AI — written for people who are new to software engineering. Three parts, about 5–6 hours for the guided work, plus an open-ended Stage 9, and you finish having built a working agent with an eval suite that scores it.
No account. No ads. Anonymous page counts only. MIT licensed.
So what is agentic engineering?
Every program is a list of steps. The only question is who chooses the next one, and when.
You write the steps in advance. The computer follows them exactly, every time — fast, predictable, and blind to anything you did not think of.
You describe the goal and let the model choose the next step while the program is running. Flexible, surprising, and it has to be checked.
Almost every real system is a mixture of the two. Choosing the mixture is the whole job — and it is what you will practise here.
Your learning path
Each part asks a little more of you than the last. Start at the top.
The Lab
Four hands-on stages in this browser. Make a real model call, extend a rule-based till until it breaks, write your own prompt, then score it against twenty real cases and watch the number move.
Open the lab→Build an Agent
Nine guided stages (0–8), then a Stage 9 transfer project on your own machine. Build the model call, rules, prompt, evals, context, agent loop, controls and defences, then transfer them to a domain you know.
↗ Starts with an on-site setup guide, then opens the local TypeScript course on GitHub.
Open the course→What you will be able to do
Tell which problems need a model and which do not
The most valuable judgement in the field, and the cheapest to get wrong.
Write a prompt and prove it works
Not by reading its answers — by scoring it against cases and watching a number.
Build an agent loop that survives failure
Tools break. The loop has to read the error and adapt, not crash.
Put a fence around what a model can do
Permission gates, mandatory review steps, and limits enforced in code.
Who this is for
Complete beginners
Parts 1 and 2 assume no programming at all.
Students and self-learners
The guided parts take about 5–6 hours, and you can pause between them. Stage 9 is open-ended.
Teachers
A 90-minute lesson plan, free to reuse and translate.
Where you are
Nothing yet. Your progress is saved in this browser only — no account, nothing sent anywhere.
Checklist
- ○The Handbook 0/11 sections Start
- ○The Lab
- ○Your first call
- ○The wall
- ○Write the prompt
- ○Measure it
- ○Build an Agent Your own mark. This site cannot see the course on your machine. Start
Next step Start the Handbook
Common questions
Is it really free?
Yes. The course is MIT licensed, with no account, email capture or advertising. The site keeps anonymous page-view counts without cookies or a personal profile. Part 1 is fully offline. Part 2 uses your DeepSeek key: signing up needs a phone number and a small top-up, and live calls can incur provider charges. Part 3 can run offline or use your own DeepSeek or Claude key.
Do I need to know how to program?
Not for Parts 1 and 2. Part 3 assumes you can read and write basic TypeScript — functions, objects, a for loop and await. If you cannot yet, do Parts 1 and 2 and come back. An older Python edition is preserved under legacy; TypeScript is the current course.
Which AI model does it use?
Part 1 uses scripted simulations and no provider. The browser Lab in Part 2 currently uses DeepSeek with your own key. The local TypeScript course in Part 3 supports DeepSeek, Claude or a scripted offline stand-in.
Is my API key safe?
The browser Lab keeps your key in session storage for the current tab and sends model requests directly to DeepSeek. The site does not intentionally send keys, prompts or replies to analytics. The analytics script runs on the same site, however, so it is a third-party trust boundary rather than a technical guarantee of isolation. Closing the tab forgets the local copy; it does not revoke the key at DeepSeek.