Orthogonality¶
The output of an agentic system is the resultant of two forces: model capability, plus everything you do at the harness layer. The model's force is one you cannot control, but you can observe it and anticipate where it is heading. It keeps growing, and the underlying architecture may change in fundamental ways.
Facing this force, the job of a harness engineer is not to chase it. It is to find a direction orthogonal to it — do what the model cannot replace no matter how strong it gets, so that every unit of effort compounds instead of depreciates.
Five articles, read in sequence:
| Article | One line | |
|---|---|---|
| 01 | The resultant | Two forces, one system. Where should yours point? |
| 02 | The first force: what is the model | An intuitive account of next-token prediction, evidence and debate on emergent behavior, and why engineers do not need to wait for the philosophers. |
| 03 | How strong is this force, and is it still growing | 2020–2026: a record of capability dimensions stacking up. |
| 04 | Is the force changing direction | Three paths beyond the Transformer: SSM, EBM, world models. How they encode the world, and what it means for harness engineering. |
| 05 | Orthogonal decomposition: where should your force point | Force decomposition from physics, mapped to engineering strategy. A criterion for telling depreciating investments from compounding ones. |
Read 01 through 05 in order. Each assumes you have read what came before.
Prerequisites: none. This is the first mental models topic.