Three systems. One boundary.Each one finds its own orbit.
THIS IS DYNAMIC SYSTEMS THEORY. Nobody tells the orbits where to go. Set the constraints, and each system organizes itself.
People · Markets · Machines
Dynamic systems theory
Movement science found that you can't program a body's every move. You shape the constraints — the body, the setting, the task — and the movement organizes itself. I trained in that idea. I've built teams, markets and agent systems on it since.
Newell, 1986: organism, environment and task constraints. Coaches call it the constraints-led approach.
The idea, in one dial
Fitted
The boundary holds. Inside it, the system finds its own pattern.
Drag it. The same dial applies to a new hire, a language model, or a sales team.
For the performer · a person, a model, an agent
Fit the boundary to what the performer can do. Widen it as they prove themselves.
Example: my Slack agent started out-of-office only, then after hours, now always on. Each step came after the team used the last one well.
For the consequence
A set point, a check, a correction. Physiologists call it homeostasis. It's how the body stays alive while everything in it moves.
Example: every weekday morning the platform checks each app against what should be running, and restarts what has stopped.
The same idea, in software · nine layers
What I do now · Tam To Target
Each client gets its own isolated cell. Each engineer gets one key. Agents work inside those boundaries, and the platform checks every morning that what must stay true still is.
Held · read from the live system, 19 Oct 2026
I test it
The same written guide helped models that couldn't do the task, and made the capable ones worse.
About 26,000 scored calls, averaged over six open models. When a model wrote the guide, capable models dropped 8.0.Once the right tool was the easy one, agents switched to it within two weeks.
Share of calls through the new CLI in its best week; 60% the week after. The written instructions changed the same week, so they get part of the credit.Guardrails a newer model could no longer even trigger. As capability grew, the constraints could shrink.
60-run benchmark at a fixed model, with control tasks.