no framework, no six-month course

Agent engineering, from prompt to graph.

Five layers, plain language, templates you copy. You already know the first one.

One payment, lifetime access. No subscription.
one unit of work, all five layers
promptthe message contextwhat it sees harnessthe machinery loopit improves graphit coordinates promptthe message contextwhat it sees harnessthe machinery loopit improves graphit coordinates
how it works

Read it in an evening, rebuild your setup the same week.

No cohort, no schedule, no calls. Written to be copied from, not watched.

1

Find your layer

Every layer opens with what breaks without it. You will recognize your own trace in one of them inside a few minutes.

2

Take the template

Each one ships with the config, the role prompts, and the check that decides pass or fail.

3

Wire the graph

The last layer puts it together: what runs in parallel, what waits, and where the work comes back.

the structure

Each layer fixes what the one below cannot.

They stack, they do not replace each other. A perfect prompt inside an empty context is useless, and a perfect context without a loop survives exactly one run.

LAYER 1

Prompt engineering

improves the message

State the task so the model stops filling gaps on its own. Role, constraints, the definition of done, the shape of the output.

Without it: the model guesses what you meant, and every run comes back different.
LAYER 2

Context engineering

controls what the model sees

What enters the window, in what order, and what gets cut first when it overflows. Retrieval, history compaction, one context per agent.

Without it: your agent reads the same file four times in one run and never knows it read it before.
LAYER 3

Harness engineering

builds the machinery around the call

Tools, permissions, response parsing, retries, limits, traces. Everything that turns a model call into a predictable unit of work.

Without it: the demo runs, production dies on the first odd response, and the log does not say where.
LAYER 4

Loop engineering

makes the work improve through feedback

Run, check, correct, run again. What counts as green, what goes back, how many attempts are allowed, and what carries over between them.

Without it: you are the reviewer, fixing the same class of mistake by hand every day.
LAYER 5

Graph engineering

coordinates all the work

Who runs in parallel, who waits, and who never runs at all. Back edges, verification gates, and the one place a human still belongs.

Without it: twenty agents run at once and the run is still sequential, at twenty times the cost.
what's inside

Templates and a rollout order, not a retelling of the docs.

Built from the breakdowns I publish, expanded to the state where you can actually run them. What fits in one line in a post is a full diagram here.

Every layer, broken down

What it is, what breaks without it, and how to tell your problem lives on this level. From real traces, not from tutorials.

Ready graph shapes

Six agents from one prompt to a finished PR. Five searchers into one sourced report. A codebase audit in twenty minutes.

Rollout order

Where to start if something already works, and what to tighten next so you are not rewriting everything at each step.

Templates

Config files, role prompts for agents inside a graph, pass and fail criteria for gates. Copy, swap in your own, run.

faq

The honest part.

Do I need this if I do not write code?

Yes, and layers one and two are usually where non-coders gain the most. Most of the material is about how to think about the work and how to shape what the model sees, which pays off in any tool you already use. Layers four and five go further into automation, and they are written so you can follow them without a background in engineering.

How is this different from free material?

Free material tells you what exists. This tells you what to do, in what order, and what to skip. Every layer comes with worked examples, the exact configs, and graph shapes you can copy straight into your own setup, so you spend your time building instead of collecting tabs. That ordering is the part nobody hands you for free.

How do I get the course after paying?

The course lives on this site, at /course.html. Access is tied to your email: you enter it there, a six-digit code arrives, and everything opens. No password to remember, and you can come back any time from any device. Nothing to download, nothing to install, and updates appear in the same place.

How does payment work?

Press get access and the payment options open on the next step. Card goes through Stripe, and takes Visa, Mastercard, Amex, Apple Pay and Link. Whatever email you enter at checkout is the one the course opens on, so nothing else is needed from you. For USDT you send me the transaction hash in a DM instead. Either way access lands the same day, usually within the hour.

Is there a refund?

The material is digital, so there is no refund once access is open. One exception: if the payment clears and access never arrives, you get everything back.

What will I actually walk away with?

A clear map of the five layers, the templates and configs behind each one, and three complete graph shapes you can run. Even if you only apply the first two layers, you will get more out of the same model with less rewriting, and the rest is there when you are ready for it.

A prompt is one line. The rest is shape.

Five layers, the templates, and the order to roll them out.

Get access