AI agents can now write code, run tests and review each other's work. What does that change for the cost of an idea, and what does it mean for your project? Part 1 of our series on the agentic way of working, with a one-week experiment and its limits.
Part 1 of 4 in a series on the agentic way of working, for teams that consider agentic AI for their own software.
For seventy years, a person typed every line of software. The cost of an idea grew with the hours it took to build it. So many ideas never got a real test, because the test cost too much.
That is changing now that AI agents can write code, run the tests and review each other's work. People still decide what to build, but the step from an idea to a working version gets much shorter.
We call this the agentic way of working. For a client, it brings two things: faster validation of ideas, and more iterations for each euro of budget.
Before the car, how far you could travel depended on muscle: your own, or the horses of a coachman. The car replaced the muscle with an engine. The coachman became a driver and reached ten times further.
A car has never known where you want to go, though, and neither does an agent. Agents take over the building. Choosing what to build stays with people, and so does judging whether the result is good, because the end user is a person too. Andrej Karpathy put it this way in 2023: "The hottest new programming language is English." You describe the destination, and the agent does the driving.
Since the end of October 2026, I have run a small agent pipeline on my own projects. It has four steps:
In seven days, the pipeline merged 134 changes. None had to be rolled back. More than a third merged between 11 pm and 7 am, while I slept.
It has limits too. It is one person, on my own projects, for about two weeks. No client code goes through it. It runs on a laptop, so it stops when I close the lid, and it runs out of memory every few days. We do not know yet how it behaves for a team, on an older codebase, or under a client's security rules. Those are the next tests.
Two weeks of data from one person prove little on their own. The experiment does show where the value of the agentic way of working sits.
This field changes every few months, and a practice that worked in spring can be out of date by autumn. So we do not publish two-year AI roadmaps, because in this field they would be fiction. We plan in quarters. Each new practice runs on our own work first, and we measure it before we keep it, so our clients get the lessons without the failed experiments.
Part 2 is about one of those failures. We built shared agent tools for our own team, and the team did not use them. You will learn why tools alone do not keep knowledge in a team, and what the second version changes.
Parts 3 and 4 cover the step from prototype to production, and the skills a team needs to direct agents well.
If you want to know what the agentic way of working could change in your team, talk to us.
