Hey there,
Last time I wrote about having my hands back in the work, building our agents myself instead of handing the job down, and about how an AI gives you your own way of working back. This carries on from that, but it is about the shape of the thing rather than the tone of it.
Something changed for me this week. I started using Claude Code on my phone. So now I walk, and I talk to it, and it builds. The dog gets me out several times a day anyway, and we moved recently so there is a lot of nature around the house. Walking in nature really does help me think better. What is new is that I can act on the thinking out there, instead of carrying it home. A thought that turns up two kilometres from a desk has to survive the walk back, and most of them do not. Now they do not have to.
A Short Walk After the Painting Class
Last night I had my painting class. It is not a class where somebody teaches you from the front. It is an open workshop, so you bring what you are working on and get on with it. I spent the last few sessions reading up on the golden spiral and the Fibonacci sequence, which is my own interest rather than anything on a syllabus. They are one thing seen from two sides. Every number in the sequence is the sum of the two before it, and drawn out, that proportion gives you the spiral. It sits in a shell, in a pine cone, in the seed head of a sunflower, and in the way a lot of the paintings I love are laid out. I want to work that beauty into my own paintings deliberately, so I was learning how it is done.
Afterwards I took the dog out, which is when these things truly start to sink in. What I was turning over was not only the paintings. The other thing I make is software. We build AI agents that do the slow work of cybersecurity compliance, and underneath them there is a library of skills and reference documents that tells each agent what the work actually demands. I wanted to know how I could work the same beauty into that on purpose. So I started looking at the library, out loud, while the dog got on with the walk.
Why the Shape Is There
So I had gone out wanting to put something in, and found it already there. Our library is arranged in a way nobody designed on purpose, which leaves the better question. Why is it there?
The answer is a limit. A model can only hold so much at once, so the knowledge has to be packed to give the most clarity per word it carries.
Which is the same problem a sunflower has, and here my evening came back round to me. A seed head has a fixed area and a great many seeds, and every one needs light. The flower solves it with the exact thing I had been reading about. Each seed goes in at the golden angle, which comes out of the same proportion as the spiral I had been drawing, and if you count the spirals in a seed head you get Fibonacci numbers. None of it is approximate. It is precise, and it is the best packing of a fixed space there is.
We did not borrow that from the flower. We came up against a hard limit and were handed a design that was already sitting in it.
I am not going to pretend I read that as coincidence. I think the design was there first. The sunflower did not invent it and we certainly did not. It holds whether you are packing seeds into a flower head or knowledge into a machine. The limit is not what makes the shape. The limit is what makes you look until you find it.
One Rule at Every Size
We keep a style guide for how all of that gets built, and it opens with three principles I have read many times. Every piece of knowledge lives in one place, and everything else points at it instead of repeating it. Nothing gets loaded into an agent's attention until a decision needs it. Each file has one job.
On the walk those three stopped being three. They are one rule said at different sizes. There is one core that holds the truth. There is one thing above it that hands out only what the next decision needs. And around those sit the smaller pieces, which point back at the core and never copy it. I checked it at every size and it held: the whole repository, then a single plugin with one gateway skill that everything else goes through, then a skill, down to one reference document with a short summary at the top telling you whether to read the rest. That is not tidiness. It is why a change only has to happen in one place, and why an agent can read a hundred words to make a decision instead of ten thousand.
Once you can see the design you can work with it instead of arriving at it by accident. So we wrote down what the flower is doing, and those same three principles came back as three things we now check.
The first is placement. The golden angle puts each new seed where it shadows none of the ones already there. Most coverage, least overlap, under scarce light. Ours works the same way. Before a new skill goes in, we look at what the existing ones cover and put the new one where the overlap is smallest. If it overlaps too much it does not become a new skill at all. It gets merged into the one it was overlapping. That takes a principle everybody nods at and makes it something you can actually check in a review.
The second is that anything new gets built out of what is already there. A new capability has to name the one or two existing ones it is the sum of. Our supplier exit planning came straight out of the continuity work and the risk quantification work already in the library. If something cannot name any parents, it is usually too early.
The third is my favourite, because it is about the person reading. They should only have to open one new layer to make one decision. If a decision needs three documents open at once, the structure is knotted and we flatten it.
What We Keep Finding
Months ago I came home from a concert hall and wrote about that evening here. The programme was about Gödel, Escher, Bach, and what had caught me was the idea of systems that watch themselves work. Not software that follows rules, but software with a layer above it asking whether the rules are catching the right things, and another above that deciding which changes are safe to make. Systems that get better at getting better, and that pull themselves back into true when something has drifted. I wrote then that it was a long way off. It was.
We keep finding that pattern one place at a time, and every time it turns out to have been waiting there already. We are being shown where things were always meant to fit.
There is a practical reason this landed when it did. The last three weeks of the cycle have been the most productive stretch of work I have ever had. I have built something like ninety per cent of the agents our product needs, so there is a fleet of them now, and the question in front of me was how they all fit together. A few weeks earlier this would have been an interesting observation and little else.
So the first thing I did with the rules was take them back through everything we had already built. Months of older material, a good deal of it written twice in slightly different words. I pruned it against the three rules, merged what overlapped, moved what was sitting in the wrong corner. It works because nothing was ever duplicated, so the whole library can be brought back into true without breaking what depends on it. What was a pile of separate agents is one orchestration now, and I am genuinely impressed by what came out the other side.
That much was directed. Me, rules in hand, going through it myself. But we have since built the checking into the system that holds all of this, so it measures its own structure against these rules and repairs what has slipped without me walking through it. That is not the whole of what I was carrying out of that concert hall. It is one step towards it. Not the product watching itself work, which is where I had been looking. The knowledge underneath it, watching its own shape. It runs on the same beauty those three were each working with in their own art, a proof, a drawing, a piece of music, and it reached me through mine. A few evenings with a spiral I wanted to paint, and a walk I would have taken anyway, with the phone in my pocket.
With care, Ben