Watched subsystems
The model is not your bottleneck. The system around it is. Learn to build the environment, state, verification, and control layers that turn a capable agent into a dependable one. You finish with a written harness in your own repository: context files, a verification gate, and a scope contract. For engineers who already run a coding agent on real code.
Reading every lesson is free; the doing layer is from $39 one-time.
Every course is included in every paid tier — today's and the ones that ship later.
Make AI coding agents actually reliable.
For you if you use a coding agent most days and still fix the same failures by hand.
Updated 2026-10-09
Build an agent that remembers, acts for you, and can be trusted.
For you if you want one agent that works on your behalf between sessions, not only while you watch it.
Updated 2026-09-27
Four shapes cover almost every agent graph you'll build.
For you if a single agent loop can no longer hold your task and you need to split the work across agents.
After: Harness Engineering
Updated 2026-09-16
Keep the loop's own measuring stick honest while it runs unattended.
For you if a coding-agent loop already runs for you and you need it to run overnight without quietly gaming its own score, burning its budget, or drifting off its real goal.
After: Harness Engineering
Updated 2026-10-09
Run many coding agents without drowning in their output.
For you if one coding agent already runs reliably for you and you want several at once, on and off your own machine.
After: Harness Engineering
Updated 2026-09-27
Run an agent-driven delivery line that earns its autonomy one gated lane at a time.
For you if a coding agent already opens pull requests for you and the question is how many of them may merge without a person reading every line.
After: Agent Fleets · Loop Engineering
Updated 2026-10-29
Design, cost, secure and prove AI systems a client can sign off on.
For you if you design, cost or defend AI systems that a client or sponsor has to sign off on.
Updated 2026-09-27
The uncomfortable part is that the failure is measurable, and it is not the model's.
What developers felt
20% faster
Experienced open-source developers, after a randomised trial using AI tools on their own repositories, estimated the tools had sped them up.
What was measured
19% slower
The same developers, on the same tasks, were actually slower with the tools than without them — and did not notice.
Our reading: this is not a model problem. It is a harness problem: unclear scope, missing conventions, no verification the agent can fail, and state that dies with the session. Each of those is something this course teaches you to fix. That conclusion is ours, not the study's.
Source: METR, randomised controlled trial of experienced open-source developers (2025). Effect sizes are from that study; your mileage on your own repository is exactly what the labs measure.
Not a mindset. Five concrete files, each from a lesson you can open.
| Before | After | File you write |
|---|---|---|
| You re-explain the project every session | The repository explains itself; the agent initializes before it touches code | Lesson 5A repository map the agent can load: what exists, where it lives, what is authoritative. |
| The agent says done; you find out later it wasn't | Done is a command that either exits zero or doesn't | Lesson 15A definition-of-done the agent must satisfy with command output, not prose. |
| One task quietly becomes four unrelated changes | Scope is a written surface with explicit exclusions | Lesson 9A next-task file with an explicit scope surface and exclusions. |
| The same class of bug returns next sprint | Each failure becomes a rule, so the class is gone | Lesson 26An incident-to-rule pipeline: the last bug you hit, converted into a rule that prevents its class. |
| Long tasks fall apart when the context resets | State lives on disk; a restart costs nothing | Lesson 11A continuity file that survives a context reset with no information loss. |
You always know what you are about to spend your evening on, and what you will be holding at the end of it.
The lesson opens on something you have already had go wrong.
Why it happens, and the smallest system change that removes it.
Questions inline, answers hidden until you commit to one.
A named artifact you add to your repository before moving on.
A concise path when you need the rule, a detailed path when you need the reasoning. Same lecture, you pick the depth.
Every file, rule, and command is written out in full inside the lesson. Nothing is behind a repository you have to go clone.
Completed lessons enter a spaced-review queue, so what you learned in week one is still there in month three.
This is a systems course for people who already write software. It is not an introduction to AI.
If you are somewhere in between — you have used an agent a handful of times and it went badly — start with the free lesson. It costs nothing to find out.
A fair comparison of the three real options, including the free one.
| What you get | Blog posts & threads | Figure it out yourself | This platform |
|---|---|---|---|
| Techniques | |||
| Ordered so each builds on the last | — | • | |
| A named artifact per lesson | — | — | |
| Project labs on your own repo | — | — | |
| Tracked progress + spaced review | — | — | |
| A certificate per course | — | — | |
| Template pack, copy-ready | — | • | |
| Kept current without re-buying | • | — | |
| Cost | Free | Weeks of your time | One payment |
Blog posts are genuinely useful and several are cited inside the lectures. What they cannot give you is order, a deliverable per step, or something that checks your work.
Not a subscription. Not per course. The platform adds courses and lectures over time, and your purchase follows them. From $39 one-time.
The curriculum is the write-up of a harness in continuous production use — including the one that maintains this platform.
Software engineer who runs fleets of self-improving coding agents across production repositories.
Noam started where you are: a heavy user of AI coding agents, fixing the same failures by hand. Today he runs a multi-role agent team that ships real product code every day, and he built this academy with the same loops it teaches. These lessons pass on what he learned the hard way: not just the tools, but the thought process behind each decision.
Every technique here was written down because it fixed something that was actually broken. The platform you are reading runs its own harness: a scheduled agent loop with a role rotation, a written contract, a verification gate it cannot skip, and a state file it hands to the next run. When that loop fails, the failure becomes a rule — and often a lecture.
That is the entire argument of the course, applied to the course. Watch the live loop.
Student results
Reserved for real, attributed outcomes from students — nothing invented, nothing placed here until someone has actually said it. If the course changed how your team ships, we would like to quote you.
Ordered by the reasons people actually hesitate.
You need to have used one. The courses assume you have run Claude Code, Codex, or a similar CLI agent on a real repository and been disappointed at least once. They do not assume you know what a harness is.
Self-paced. Harness Engineering is about 10 hours of reading across 23 lessons; Agent Systems is about 3 hours of reading across 8 lessons; Graph Engineering is about 2 hours of reading across 10 lessons; Loop Engineering is about 3 hours of reading across 8 lessons; Agent Fleets is about 7 hours of reading across 16 lessons; Software Factory is about 14 hours of reading across 33 lessons; AI Solutions Architect is about 9 hours of reading across 15 lessons. The project labs are real work on your own repositories, so they take as long as your agent takes.
30-day money-back guarantee, no questions asked, no form to fill in. Email and it's refunded.
Yes. One payment, lifetime access — including every future lecture and every future course on the platform. No subscription. Multi-seat billing (Team) is in development.
The learning system around the material: structure and sequencing, tracked progress, concise and detailed paths per lecture, self-checks, spaced review, labs on 7 of 7 courses, a certificate, the copy-ready template pack, and lifetime updates.
Reading. Every lesson of every course is free to read, no card required. Paying unlocks the doing layer — saved progress, recorded quiz answers, verified completion, spaced review, labs on 7 of 7 courses, and the certificate — from $39 one-time.
The material is agent-agnostic: it teaches the environment, state, verification, and control systems any CLI coding agent needs. Examples and labs use Claude Code and Codex; the techniques apply to Cursor, Aider, and whatever ships next.
Bring-your-own-key. Rubric labs are scored inside your own agent, on your own Claude or OpenAI key; checker labs run a local script with no AI at all — your code and API keys never touch our servers. Pro and Team also bank grading credits for server-side grading, which is in development — credits are granted now and spendable when it ships.
The model names will be. The systems will not — that is the whole argument for teaching the harness rather than the tool. Updates are included forever, and the curriculum is edited continuously rather than re-recorded every year.
Every lesson is free to read, no card needed. Start with the first one. If it does not change how you brief your agent tomorrow, nothing else here will either.