The seven principles, and where each one breaks

Where it came from

I joined a project that had already been running for a while. Stores were built, e-commerce was live, and the team was about to buy an ERP. My instinct said no. Should we really have an ERP, or is this the thing we build. The advantage we had was going to come from an agent driven stack of our own, buying only the pieces we genuinely needed. I could only have that argument because I was in the stack myself. You cannot form that view from a status report. Technology shapes the organisation, and the organisation shapes the technology back.

When it does not hold

I do not think being close to the technology ever cost us anything. The cost sat next to it. I change my mind easily, and we moved several foundations more than once, the LLM layer and the development environment among them. A leader who is deep in the stack is also a leader with opinions about the stack, and it is the team that carries the churn. Worth knowing before you copy this.

Where it came from

Vendors pitched us everything, SEO tooling, staff planning, all of it. Most of it was work we were better off doing ourselves, partly because the people who knew the trade were in the room specifying it, and partly because of what a SaaS contract actually is. You buy ten percent of what you want, and ninety percent that somebody built for a need that is not yours. In a small team nobody has time to learn a large system anyway. It has to simply work. We built the goods supply flow ourselves, and an agentic PIM.

You buy ten percent of what you want, and ninety percent that somebody built for a need that is not yours.

When it does not hold

We bought Shopify, an email system, and a checkout with payment and delivery selection.2 The rule was simple. Where our own competence was thin enough that we could not specify what we needed, we bought. Where the time to build outweighed the time to market, we bought. In hindsight I would have moved off Shopify to Medusa earlier for the control it would have given us. Buying is right more often than build-everything people admit.

Where it came from

Change management can happen over a lunch if everyone is at the table. There are no other dependencies, which is what makes a small team fast rather than merely cheap. When we decided to move our messaging from Teams to Slack, it was decided in one meeting and carried out more or less the same day, and everyone understood why. The other thing smallness buys is the view. Each person at that table has the whole picture, because they talk to everyone, and depth in their own area at the same time.

Change management can happen over a lunch if everyone is at the table.

When it does not hold

Being few means you cannot do everything. Nobody gets more than the hours they have, and an agent still needs a person somewhere in the loop. I think that limit was good for us. It forced us to choose. It also set the selection rule: we dropped anything that could not grow with an agentic approach, which is a narrower filter than it sounds, and it means a small team is not a neutral choice. It decides what the business can become.

Where it came from

The full account of the plateau is in issue 02. The short version is that everyone got good at the easy layer of the tools, and then stopped. What moved people up a level was not a decision I took. We ran a hackathon every week, two or three hours with the whole team, talking about what we were building and coaching each other. People saw what the rest of us were building, we worked out the next step together, and I pushed a bit. I cannot point to the exact moment it turned.

When it does not hold

There isn't one, and I was askedI was asked when it is right to let a plateau stand, and I do not have a case. I do not think you should stay on the ledge. That is a position rather than a limit, and I have written it as one.3 The nearest thing to a caveat is that pushing has a cost, and the person you push has usually just got comfortable with something that was hard last month.

Where it came from

The whole team were domain experts apart from one technical hire, and the two of us built around them. Deep domain knowledge lets you skip an enormous amount of research. Launch windows in Swedish retail, how suppliers and transport actually behave, none of that comes out of an LLM cleanly. Building a store is not something you do with AI. You need a carpenter, and you need someone who has done it before and knows that. The people who moved fastest were the ones willing to write down how they think.

When it does not hold

There isn't one, so here is the pairing insteadI was asked for an occasion where domain knowledge ran out and someone got stuck, and I do not have one. So I will state the pairing instead. The technical half of this is real, and it is a third skill rather than a version of the other two: deep in the LLM world rather than in systems architecture. Neither half worked alone.

Where it came from

We tried moving off Teams once and failed.4 We announced it and people simply stayed where they were. The second attempt stuck, because by then we had found that what we wanted to build could not be built inside the platform we had, and because Slack is made for developers and Teams is not. The thing I remember better is a hackathon where I apologised in advance for the next switch, away from Claude, at some point in the future. I watched people shift in their seats.

When it does not hold

The limit is not in the principle, it is in who pays. We should have left Teams much earlier, and we should have moved from Shopify to Medusa earlier, so our failure was staying too long rather than switching too soon. But that apology in advance exists for a reason. Every switch takes something away from people who have just got good at the thing you are removing.

Where it came from

This one has been rewritten, because the version I published first claimed more than I can stand behind. There was no day when a machine made a call it should not have. The shape was chosen in advance instead. Nearly everything we built was a combination of agents and ordinary programmatic code, with a person somewhere in the loop, and that combination is what worked. Agents where the work repeats, code where the rules are known, people where context and judgement live.

When it does not hold

The line moves, and mine moved fast. In the last stretch we were working on agentic dynamic pricing and campaign planning, and I was ready to hand both to a hybrid of agents and code. More of our communication work could have been automated than we managed. So this may be a description of how far we got rather than a law. It is also possible it stays true for longer than people expect.