Stage 3 · How it gets to work
Harness
What surrounds the model?
All the software around the model that makes it usable: the interface, system instructions, memory, tools, permissions and the loop that decides when to stop. ChatGPT, Claude Code, Codex and Cursor are harnesses.
An image to remember it
If the model is the engine, the harness is the car: steering wheel, brakes, dashboard and seatbelt.
The misunderstanding
If I switch apps, I switch AI. Sometimes it's the same model in a different harness, and the difference you notice comes from the harness.
What changes in practice
When something doesn't work, ask whether it's the model or the harness: did it have access to the file? Could it search? What instructions did it come with out of the box?
