Agentic Engineering
Today a fleet of AI agents builds, tests and ships enterprise-grade software for my companies and for clients. At a fraction of what a classic development team costs.
The backstory
Where most of the 1,000 hours happened
As an SEO agency we suddenly found ourselves in a market where clients cancelled. Everything was supposed to be solved with AI, and many clients wanted to do things themselves. To stay competitive and to open up new lines of business, I decided to go into software and to learn everything there is to learn about it.
So I hired several full-time developers. Full-stack engineers with 15 years of professional experience. For more than twelve months we tried to build one piece of software together. It went slowly. They did use AI, but not the way I knew it should be used: a bit of code completion here and there, and a great deal of work by hand.
When I described the way of working and the speed I had in mind, I was told it could not be done like that.
I saw it differently. Over the following twelve months I put hundreds of hours, about a thousand in total, into research and hands-on practice to answer one question: is that true?
It is not. There are ways to do it. And they rest on software architecture principles that have been sitting in old books for decades.
How a run looks
I tried practically every agentic development system on the market. I stayed with a setup of my own. It is not the perfect setup for everyone. It is the perfect setup for what I want: enterprise-grade software, at speed, without giving up quality control.
Simplified illustration of a typical run. Tool and model names shortened.
Principles
None of this is magic. It is classic software engineering, enforced by machines instead of hoped for from people.
A module is deep when its interface is much smaller than what it hides. Agents work well inside clear borders and badly in a tangle. The idea comes from John Ousterhout’s “A Philosophy of Software Design”, and before that from David Parnas in 1972.
Every module is tested through its public interface, the way a user or another system would call it. For critical modules the tests exist before the code. An agent cannot talk its way past a failing test.
Speed does not come from skipping quality control. It comes from automating quality control and placing it in front of the AI as a guardrail it cannot bypass: static analysis, architecture rules, CI.
I work with the frontier models of Anthropic, OpenAI and Google at the same time. One lab’s model builds, another lab’s model reviews, a third cross-checks every critical finding. They catch each other’s mistakes.
A dispatcher splits work into independent packages and starts a worker for each one. On large jobs hundreds of agents contribute. Nothing merges unless every gate is green.
All agents read from one shared memory and follow one versioned set of rules. What one agent learns on Monday, every agent knows on Tuesday.
A guard blocks catastrophic commands before they run. A gatekeeper checks every piece of external content for prompt injection before an agent may read it.
Staging runs fully autonomous. The step into production belongs to a human. Autonomy is a dial per module, not a switch for the company.
Standing on shoulders
I taught myself, but not alone. Two people shaped my setup more than anyone else: Matt Pocock, who shows how deep modules and test-driven work make agents reliable, and Kun Chen, whose open tools for orchestrating and supervising agents I use every day. The foundation under both is older than any language model: Ousterhout, Parnas, Fowler, Beck.
The field moves fast. A workflow that worked a month ago may be obsolete today. I replace mine constantly, and I plan to keep doing that.
What we built
Before I sell anything, I use it. We built practically everything our agency needs: more than 40 interconnected systems and services, all talking to each other.
An SEO suite, a content factory that writes and publishes by fixed rules, a monthly radar across every client, pages built for AI search, and a pipeline that turns text into finished WordPress pages.
Monthly bookkeeping preparation, receipt recognition with clean exports for the tax advisor, payroll distribution and invoicing. Mostly without a human touching a file.
Deep-dive audits with nine agents in parallel, legal research that keeps itself current, market and trend monitors, scanners that alert me when something matches my criteria.
Our own e-signature, a coach that reviews every sales call and sends tips, an AI role-play trainer for phone and closing calls, verified lead imports.
Time tracking and budgets, an access portal, a task board, a knowledge base, a whiteboard, offline speech-to-text, an inbox that sorts itself. They replaced a stack of subscriptions.
Automatic deployments after green checks, monitoring, consent management, the guard and the gatekeeper. A double-digit number of servers with dedicated cores, mostly in Germany, some elsewhere in Europe and in the US.
Virtual employees
On top of the systems we built agents with a name, a job and a schedule. The team writes to them like to any other colleague.
SEO
Crawls every client domain each week, watches what competitors publish, runs audits, finds topics and writes articles.
Finance
Finds and creates invoices, prepares reminders, pulls time-tracking data and sends letters.
Google Ads
Audits accounts and plans campaigns against a fixed rule set. His knowledge base holds 15 books and about 2,150 sources.
Client work
Listings, enquiries, linked email and publishing to the major property portals. Specification first, tests first, autonomy levels per module. More than 1,700 automated tests guard the application.
An agent turns every new request into a drafted task. A human approves it with one click.
A website relaunch with an automated publishing pipeline, a tariff directory maintained by an agent, and a topic radar tuned to finance and insurance.
Tell me what it should do. I will tell you honestly whether my setup is the right one for it.