★ FEATUREDBuild your BI dashboard from scratch — faster than taming Power BIRead →
Solutions · Agentic Engineering

We don't write the code.
We built the machinery that does.

34 repositories, 15 flagships and the 123 pages you are standing on were written by AI agents. Not as an experiment — these are systems running in production for paying clients. Christian has not written a line of them. He decides what gets built, and when it is good enough.

Book a meeting →See the flagships →
34
repositories in production
8 weeks
to an entire health platform
1995
building since

Getting an AI to write code is not the hard part. Daring to ship it is.

Anyone can get a model to produce a hundred lines that look right. It is free, it takes thirty seconds, and it is not what the work consists of. The work is knowing whether those hundred lines actually work, whether they broke something else, and whether anyone would notice if they did.

So the interesting part is not the model. It is the machinery around it — the gate that stops a release when a test goes red, the card that has to be written before the code, the rule that a result must be proven rather than claimed. Without that you build something fast that nobody dares touch again.

We spent two years building that machinery for ourselves, because we needed it. That is what this page is about — not about prompting better.

The loop

How a task goes from idea to live.

1
The idea becomes a card

No agent starts coding from a conversation. The task is written down with acceptance criteria — what must be true before it is finished. That is what makes the work judgeable afterwards instead of merely believed.

2
The agent builds — and judges its own work

An AI agent picks up the card, reads the repository and builds. Then tests, type checks, a security scan and a visual check in a real browser. If one of them goes red, nothing ships.

3
The human curates

Christian looks at the result and says yes, no, or do it again. That is where taste comes in, and it is the only place a human is required.

The principles

Six things that make the difference — learned on our own mistakes.

Context is the product

An agent is only as good as what it can read. We spend more time writing down how a system fits together than writing code. That is where the quality comes from.

The card before the code

No agent invents what to build. The task and its acceptance criteria are written down first — otherwise something nobody asked for gets built, and nobody notices until it is there.

Proof, not claims

«It works» is not a result. A screenshot, a test that goes red when you break something, a value read back from the database — that is a result. We have measured our own agents report green on things that were broken.

The gate is not negotiable

A red test stops the release. No agent may skip it, and none may fix the test instead of the fault. A gate you can talk your way past is not a gate.

Reuse beats re-rolling

49 shared packages mean mail, login, AI access and design tokens live in one place. A copy is not wrong the day it is written — it is wrong the day one of them gets fixed.

Delegate, don't dictate

You do not tell an agent which lines to write. You tell it what must be true afterwards, and let it find the way. It is the hardest habit to learn and the one that pays back most.

The proof

We don't recommend anything we don't run on ourselves.

The fleet · 15 flagships
The tooling is ours, and it is in production.

cardmem holds the plans and the board. lens looks at the screen in a real browser and catches what tests cannot. trail remembers why a choice was made, so it never has to be derived twice. buddy ties the sessions together. They are not demos — they are what we build with every day, and they were built by the agents they govern.

Case · fysiodanmark
Eight weeks to an entire health platform.
Platform

Booking, staff, care pathways, PWA and operations — built and shipped in eight weeks. Not because anyone typed faster. Because the work could run in several tracks at once, and because the machinery caught the faults before the client did.

Honestly: getting here is not free. The first months went on building gates and learning to write down what you mean. The payoff comes afterwards — but it comes.

Would you rather learn it yourselves?

We run a workshop at your place.

Two days with your own developers in your own repository — not a course example. We build something real together, set up the gates, and you keep the documentation. Tool-agnostic: we use what you already have.

2 daysAt your officeIn your own repositoryTool-agnosticYou keep the setup
Book a workshop →
Questions

What people usually ask.

No. Autocomplete suggests the next line while you type. An agent reads the whole repository, makes a plan, edits files, runs commands and checks for itself whether it worked — and keeps going until it does. The difference is not speed, it is who holds the task.
Because it cannot ship until it has proven it. Tests, type checks, a security scan and a visual check in a real browser run on every single change, and a red test stops the release. We have measured our own agents report green on things that were broken — which is exactly why the gates exist.
We are. There is nobody else in the loop. That is also why we spend so much effort making faults visible quickly rather than trying to avoid them entirely.
No. The models change several times a year, and the machinery is built so that can happen without touching the code. All AI access goes through one place, so a switch is one setting — not a rewrite.
Anything involving personal data runs on European models. We do not send client data to US services, and we can answer where every single request landed.
We can start in what you have. The first step is writing down how the system fits together — which is what is missing in most places anyway, and what an agent needs in order to be useful.
Agentic Engineering

Shall we build something,
or teach you how?

30 minutes with Christian. No obligation, and an honest answer — including if the answer is that you don't need us.

Book a meeting →
Hi — I'm Aidan