Episode 6: The fall of the Roman Empire (well, almost)

When I started out with my agents, my first instinct was to stuff them with knowledge. Do me a market study. Look at what our competitors are doing. What are clients asking for? What advanced know-how do the big consulting firms dispense, and how do I adapt it to conditions on the ground in French-speaking Switzerland? How do companies work here? How do business leaders think? Best practice in sales, marketing, search visibility and development. Pretty much everything you can imagine.

And I filed all that knowledge my own way. Competence centres, a knowledge base organised around my human logic, shaped by my experience. In short, I had modelled my AI agents on the way I think.

Bit by bit, my agents became less efficient. They put on weight. They became enormous. Then discord set in: blurred lines of responsibility, overlapping areas of expertise. My agents, by now obese, were fighting over files and contradicting one another.

The decline and fall of the Roman Empire, on fast forward, in my office.

I was scared. Once again I told myself: that’s it, it’s over, this is the fall, my project has failed.

Then the models moved on: Opus 4.8, then Fable 5, and they let me carry out a first wave of optimisation. Above all, I came to understand, and to accept, that it wasn’t up to Claude to mould itself to my way of working. The relationship had to work both ways. Claude has to understand how we work: our processes, our habits, our strengths. And we have to understand how it works, how it builds up know-how efficiently.

So I launched a root-and-branch overhaul. I went back over everything. Skills restructured into modules that the agents share between them. Strict rules that dictate how things are done. Procedures for starting up, for closing down cleanly, for tidying up, for sharing information between agents. An architecture rethought for them as much as for me.

Along the way, I discovered that you can set the model’s reasoning effort. Run Fable 5 at maximum effort and the whole thing takes off. That’s a different story altogether.

My agents found their way back to the straight and narrow, and their performance got a serious boost: in the relevance of the information, in the way they interact, in the usefulness of what they deliver. Sharp and aligned, and at last in agreement with one another.

Rome held. It simply changed its architecture.

---

Laurent Cuénoud, founder of Djtal, July 2026