Writing
Essays on AI, operations, and what is actually working in the field.
I write when something nags at me long enough that writing is cheaper than not writing. The throughline across most of these is the same: the AI conversation is upstream of the AI work, and most organisations are getting the order wrong.
9 essays · roughly one a month
2026
Why most enterprise AI pilots die in committee
5 min · May 2026
The pattern is so consistent it should embarrass us. The reason is not the technology. It is that pilots are being scoped, governed and reviewed the same way ERP migrations were scoped twenty years ago.
The moat is not your model. It is your ideas.
7 min · Apr 2026
Every team has access to the same models. The interesting question — the only interesting question — is what they decide to point them at, and how confidently they hold that decision.
The path of least resistance for agents
7 min · Apr 2026
If your AI worker keeps asking for clarification, the problem isn't the model. The problem is the surface you handed it. The interesting craft is shaping the workspace so the agent can just go.
Drift is the silent killer of agent swarms
8 min · Mar 2026
More agents in parallel feels like more throughput. It can be — but only if you have built the drift-prevention architecture underneath. Without it, the swarm invalidates its own work.
First principles before second principles
6 min · Feb 2026
Most agent failures are not failures of execution. They are failures of alignment upstream — clever solutions to the wrong problem. The cheapest move is to ask 'what is the actual problem' before evaluating any answer.
Mission shape: zoom out to one hundred thousand feet
6 min · Jan 2026
Without a 100,000-foot view, every local decision is rational and the aggregate trajectory is nowhere in particular. The cure is dull: write down the load-bearing claims first, and only commit work that advances one of them.
2025
The room is the architecture
7 min · Dec 2025
The interesting unit of senior knowledge work is no longer the conversation with one agent. It is the room — the conductor, the workers, the audit loop, the marketing cron. The operator becomes the architect of how agents talk to each other and to the work.
If you ask for it twice, automate it
6 min · Nov 2025
Most of the wasted cycles I see working with AI agents come from the gap between 'I noticed I am doing this often' and 'I built a thing so I do not have to do it any more'. That gap is where the entire compounding game lives.
Code that writes code
8 min · Oct 2025
The interesting questions are no longer 'what does the agent do for you'. They are 'what do you let the agent build, and how do you preserve enough understanding of the result to remain its operator'.
Follow along
Short notes, new essays, and the occasional video. Pick whichever channel you already live on.