agent-harness-building
Installation
SKILL.md
Harness Engineering
Agent = Model + Harness
The model predicts text and does nothing else. The harness is everything wrapped
around it that turns prediction into work. When the model emits read auth.py, it
read nothing — it produced text shaped like a request. Opening the file, handing
back the contents, and asking what to do next is the harness.
Every concrete action in an agent run is the harness acting on text the model only proposed.
When you actually need to build one
Most teams inherit a harness and only author a charter. Build one when the workflow you need is one no stock harness can express — not because building one sounds interesting.