Agent engineering for people who build systems without writing code
You have been using an AI coding agent for a few weeks. The first two were extraordinary. Then something changed.
It started telling you things were done that were not done. It confidently produced work that looked right and was not. You asked it to fix something and it fixed something adjacent, said it had succeeded, and moved on. You started checking everything by hand, which is roughly where you were before.
Nobody warned you about this, and the tutorials do not cover it, because the tutorials stop at the part where it feels like magic.
That ceiling is not a limit of the tool. It is the point where using an agent stops being about prompting and starts being about engineering, and nobody taught you the engineering because you were never supposed to need it.
Not how to code. You will not write code here, and neither do I.
What you will learn is the judgment that decides whether an agent system works: what to ask for, what it can see, whether it actually did the thing, how work should be shaped so agents compound instead of colliding, and what should survive between runs.
Five judgments. They are what separate someone who uses AI tools from someone who engineers with them, and right now they are what companies are hiring for.
People who build systems without writing much code. Salesforce solutions engineers and admins, ServiceNow and Workday consultants, RevOps, technical PMs, Zapier and n8n power users.
You already reason about state, triggers, environments, and data models. You have debugged an automation that reported success and changed nothing. You have argued about which of two documents is authoritative.
Nobody has told you that those instincts are the entire game in agent engineering, and that the vocabulary you are missing is smaller than you think.
Where people go from here: Forward Deployed Engineer and AI Solutions Engineer roles. Those titles barely existed three years ago and are now among the most in-demand in the industry, and the profile they want is closer to yours than to a backend engineer's.
Not notes. Six things:
The demo is the one that gets you interviews. Every student solves the same client problem, deliberately, so yours is comparable to everyone else's and worth putting in front of somebody.
Five units, four to six weeks, self-paced. Roughly ten hours a week.
Your machine runs nothing. You connect whichever AI you already use, and your work happens in a workspace on the server. Any MCP client will do: Claude Code, Claude Desktop, Cursor, ChatGPT, Gemini CLI.
| 1 | The Loop | An agent that runs on a schedule and survives running twice |
| 2 | Context and Proof | It can no longer lie to you |
| 3 | The Org | One agent becomes several, with a memory they share |
| 4 | Production | Real integrations, and an eval that proves quality |
| 5 | The Engagement | A client, a week, a demo |
You never write code. You write specs, context, verification criteria, and orchestration. Your agent writes the code. Then you read it, which is the actual job.
Every lesson breaks something first. You do not get told about idempotency. You run your loop twice, get two duplicate outputs, and then get five minutes on why that happens and what it is called. The concept arrives attached to a scar.
Every unit has a gate. Automated, staged, and it does not believe you. It reads what your system produced, never what your system said it produced. You always have exactly one thing to fix next.
Interview prep runs alongside from week one. Spaced-repetition drills, a mock interviewer you run yourself, and every concept comes with the scenario as an interviewer actually asks it, the follow-ups, and the wrong answer that sounds right.
That last part is worth the price on its own. "Wrap it in a transaction," "set temperature to zero," and "we should increase test coverage" are three answers that sound like competence and cost people offers.
It is built from a system that has been running in production for months, including everything that went wrong with it.
The truncation lesson is a real memory index that grew to 39KB against a 24KB read limit and silently hid everything below the cut for weeks. The verification unit is built on an agent that reported success every morning for four days and produced nothing. The concurrency unit is a real lost-update bug.
You will not find these in a tutorial, because tutorials are built from things that worked.
A job. Nobody can promise that and anybody who does is selling something else. What this promises is the artifacts and the fluency. What you do with them is yours.
That it will be comfortable. Unit 2 is designed to make you distrust your own tests. Unit 4 will probably show you that something you built is worse than you thought. That is the course working.
That you can skip Unit 0. It is a free 90-minute check that you can work a terminal. If that stops you, the rest will be miserable in a way no amount of help fixes, and it is better to find out for free.
You run agents on your own AI account for six weeks, including some unattended. Unit 0 has you set a spend cap before anything runs, so there is a ceiling you choose.
A measured figure will go here once the first cohort has finished. Until then, ask and you will get a range with the working shown.
Founding cohort: five seats at $500.
In exchange I ask for a written account of how it went, honestly, and permission to publish your final demo. After those five, it is $1,200.
The founding price is lower because you are going first. Nobody has taken this yet, so you will hit rough edges the fifth student will not, and the price reflects that rather than pretending otherwise.
Free, ninety minutes, and a filter rather than a sales funnel. Set up your AI, run a command, edit a file at a path, set an environment variable that persists, and set a spend cap.
Finish it and this will work for you. If you cannot, do not buy it, and tell me which step stopped you. That is worth more to me than the sale.
Built and taught by Sully K.
a loop that runs on a schedule, does one real job, and survives being run twice.
6 lessons · about a weeka fail-closed verification suite on your loop. Your agent can no longer lie to you.
6 lessons · about a weeka supervisor, isolated workers, and a memory they share.
6 lessons · about a week and a half. The heaviest unit.two live integrations and a scored eval suite.
9 lessons · about a weeka recorded demo, a writeup, and a running system solving somebody else's problem.
6 lessons · one to two weeksEach unit ends with a gate that runs against what you built. It checks your artifacts rather than your agent's summary of them, and when it fails it names the one thing to fix next.
There is no video to watch. You connect whichever AI you already use to this site, and it fetches your lessons, runs your work in a workspace on the server, submits your gate results, and passes your questions to a human who answers them.
Works with Claude Code, Claude Desktop, Cursor, ChatGPT, Gemini CLI, or anything else that supports MCP. Setup is one command, about ten minutes. Your machine runs nothing.
Unit 0 is free and takes about ninety minutes. It checks you can work a terminal well enough to enjoy the rest. If it stops you, you have lost an evening rather than a fee.
Nothing to pay today. Five founding seats are open.