Thinking
An observation is cheap. A project costs money, time and other people's careers. What sits between them is the work of deciding which observations deserve the expense.
Each promotion costs more than the last, which is the only reason to separate them. A good sentence is not a validated system, and treating one as the other is how expensive mistakes start.
It does not always run in that order. The chess analogy sat as an observation for years before it explained anything. Sense, decide, act, show arrived as a framework first and had to be argued backwards. Most things stall at note, which is where most things belong.
Four stages, and the fourth is the one that gets dropped.
A system that senses, decides and acts without showing its work is one a person cannot correct, cannot audit and eventually cannot trust. Showing the work is what keeps someone's understanding intact while the labour moves away from them. It produces no visible capability, which is why it is the step that gets cut.
Data is not one of the four. It is the floor they stand on. An agent is bounded by what it can sense, remember, understand and verify, not by the model behind it.
The loop sits inside a larger one. A system regulates the people who use it through five channels: information, incentives, effort, feedback and ambiguity. Only effort is visible on a screen, which is why it gets most of the attention, and incentives and ambiguity are usually the stronger two. I have worked that argument out at length in A General Theory of Human-System Interaction.
Long-form, structured, carrying evidence through.
Why the gap between wanting an outcome and having it is hard, what an analytics product taught me about compressing it, why 53 seconds still did not close it, the difference between operating software and instructing it, and what an application becomes responsible for once nobody is driving it step by step.
Working papers. Longer, more formal, and provisional by design. Both are versioned and carry a changelog.
An eight-layer trust model with strict dependency ordering, mapped across human, agent-human and agent-agent interaction, plus how trust attenuates across delegation chains. It argues that incentive alignment is the layer nobody has solved.
Where control, context or commitment crosses between reasoning parties, things get dropped. Three seams, six property axes, ten design principles, a seam-coverage metric, and a hypothesis stated so it can be shown wrong.
Felt ease treated as a regulation problem rather than a design problem. Complexity is allocated rather than removed, and the audit question is who absorbed what was taken away.
Short, numbered, undated. Each one is a position I would defend, compressed to the point where the reasoning is still visible.
The permission a system has should track the consequence of the action, and the infrastructure required to hold it.
The shape of business software was set by constraints that are now weakening. The shape has not moved with them.
The problem we could not solve was solved by people who were not solving our problem.
Things noticed and not yet argued. These are the least defended items here, and they are the stage everything else came out of.
People say they need to buy something properly. The system asks them to create a requisition. The translation between those two sentences is where most of the friction lives, and it is performed by the human every single time.
Ask somebody which tool they trust and they will usually describe something that made a situation legible at a glance rather than something that did work for them. That may be about trust rather than about interfaces.
Demonstrations fail on capability. Deployments fail on the edges, the cost per run, and the question of who is responsible when it is wrong.
When an answer took weeks, requests were rationed and carefully chosen. When it took under a minute, the questions got smaller and more numerous. I do not yet know whether that was an improvement.