I put AI agents into production. Here's what they don't do.

I don't use Claude Code to write more code. I use it to extend what I can do as an architect: analyse, implement, test, review, document, deploy and monitor, with guardrails that keep the automation from turning into a black box.

I work independently, between Belgium and Morocco, with more than 30 years in the field, in data architecture and in IT for small and medium-sized businesses (SQL Server, C# and .NET, ETL). My starting point, drawn from experience: a problem in a small business is rarely just technical or just human. It is both, seen from two angles. AI does not change that. It changes what an architect can do for each client.

In practice, I currently run a data project for a small business, with a software product, a shared framework, several websites and an internal governance repository, while remaining solely accountable for the decisions.

Why this system? Computers have been part of my life since I was 10. Two bouts of exhaustion taught me what it costs to want to carry everything alone, and led me to coaching, with a certification in 2019. Today, the system carries the memory, the rules and the checks. The accountability stays with me. The time it gives back goes to my clients: understanding their business and delivering what works. For them, that means fixes verified before production, and a written record of every choice.

Who decides, who executes?

Everything rests on a clear boundary: I make the business decisions, the agent executes, checks and documents.

The agent can propose, implement, review and document. Accountability stays human and cannot be delegated. These rules apply to whoever executes, agent or developer.

How does a request travel from the issue to production?

Every request, whether it arrives by email, by phone or by message, first becomes a GitHub issue. Nothing gets written without an issue. Each issue receives a score, using the method described in the box: the more money at stake and the more records affected, the higher it moves up the queue.

A textbook case: an anomaly affecting 250 records in production, with money at stake, and one day of work. Value 5 × impact 5 × risk 3 × production 2, divided by an effort of 3: score 50.

How does the agent keep the memory of the portfolio?

Data project

Production: supplier data flowing into a business management system, with consistency checks.

Product

The product: a cross-reference search engine, neutral with respect to the business domain, with an API and a dashboard.

Framework

The foundation: my set of .NET libraries, reusable from one project to the next.

Branding

The websites of my businesses.

Governance

The logbook: decisions, ADRs, a journal shared by all the repositories, audit reports.

The governance repository acts as the memory: its decisions and its journal are loaded into every session. The agent never starts from scratch.

What guardrails frame the agent?

An agent's speed is only worth something if it is framed. My standing rules are read both by the agent that writes the code and by the one that reviews it. They improve with every iteration: an error that is caught becomes an instruction.

• A specification, then a ten-line design before any code, for every structural choice. Business choices come back to me, never settled quietly.
• A self-review backed by mutation testing: every fix has a test that fails without it.
• The full test suite after every fix, and a written question: what could this fix break elsewhere?
• An independent review before every merge.
• No silent errors, no guessing: a missing value is reported, never replaced by a zero or an old price.
• Test before production: my process aims to write to production only after approval on Test, and every corrective action is logged.

These rules do not make the agent infallible. They make its errors visible early.

What does a real agent error look like?

A few weeks ago, agents triaged 226 open requests and proposed that I close 21 of them. I closed none on their word. I gave three other agents the opposite mission: prove that each closure was wrong.

One of the cases: a dashboard showed the default environment instead of the one requested, without any warning. The triage said it was fixed. The checker found the defect in the code.

In all, 9 of the 21 proposed closures were refuted. The other 12 held, and I closed them. An agent's proposal is not a decision.

What do the automations do every night?

An automation starts out read-only: it observes, reports and proposes before it earns the right to act.

What does it add up to, in figures?

Period measured: 1 April to 9 October 2026.

I keep the number of lines of code out of this table: it measures volume, not value, and part of it is generated code.

AI has not reduced accountability. It has increased the capacity to steer.

If you want to put AI agents to work without losing control, let's talk: a first conversation is enough to see what carries over to your situation.