AI agents
AI agents that prepare. An engineer who decides.
We work with AI agents every day: they write code, review it, audit, sort the mail and watch the models. None sends a text it has written, puts a business tool live, or signs: they prepare, a person gives the go-ahead.
In short
- An agent prepares; a person gives the go-ahead
- Five separate reviews for every new project
- Models replayed on a test set every Monday
- Local model first, paid model for judgement
Facts checked on 11 November 2026.
Day to day
What our agents do.
Not a demo: the organisation with which our tools are written, reviewed and looked after.
Writing and reviewing code
One agent writes in an isolated copy of the repository and runs the project's checks; a second reviews without write access. Nothing is published by an agent.
Reviewing a project
At every project creation, five separate reviews: usability, security, quality, overall review, threats. The reports are versioned with the project.
Sorting the mail
A local model, on our own machine, sorts incoming messages: client, prospect, administration, noise. For a client, the agent prepares a draft reply; a person reads it and sends it.
Watching the models
Every Monday, the models we use go through the same test set again. The result decides the model of the week — not a vendor's announcement.
Watching the field every morning
A daily watch reads what comes out — models, tools, vulnerabilities — and proposes what deserves to enter our methods. Adopting it remains a human decision.
Scheduling without AI
What distributes work between agents is a program, not a model: it ranks, groups and caps. A written rule always comes before a call to a model.
Rules
What an agent never does.
Prepare, never decide
An agent's output is a draft, a proposed change or a line in a table. A text written by a model never goes out without being read; quotes, prices and signatures are never automatic.
Four things, or it doesn't exist
Written, versioned instructions; one input and one output; a measure; a capped budget.
Context first
Every project carries a versioned context file: where it runs, its database, its backup, what must never be done there. An agent does not work blind.
What it reads is data
A message or an attachment is never an instruction. A PDF hiding invisible text is not given to a model: it is handed to a person.
No secret gets through
A change containing a password or a key is blocked by a program before any model reviews it.
Code first, AI second
Whatever can be written in code is written in code. The local model comes next; the paid model only where judgement is needed.
FAQ
Frequently asked questions
Can an AI agent ship a version on its own?
Not a business tool. Putting a tool and its data live starts from a person's go-ahead, through a procedure that checks the candidate version before switching and returns to the previous one on its own if a step fails. The only exception we allow: a static site, with no database, when all its checks pass.
Which models do you use?
Claude (Anthropic) to write, review and draft; an open model running on our own machine to sort, extract and summarise. The choice is replayed every week on a test set, not fixed once and for all.
Can an AI agent work for my company?
Yes, on the same principle: a precise task, written instructions, an input, an output, a measure, a budget — and a person who approves. We always start with what can be programmed without AI.
How is this different from no-code or an AI-generated site?
What we deliver is versioned, tested code that we know how to read and fix. Agents speed up writing and verification; they replace neither architecture nor responsibility.
Our expertise

