noun · AI concept
Harness
also called agent harnessscaffold
The app wrapped around the AI. The model only talks; the harness is the part that hands it files, runs its commands and remembers things.
The application code around a model that assembles context, calls the model, executes tool calls and manages the loop and the conversation state.
The program that builds each request, sends it to the model, reads what comes back, runs any tool the model asked for, and loops. It decides what goes in the context window, so much of an agent's behavior lives here.
Real-life example
You ask a coding assistant to fix a bug. The model never touches your files. It writes back "read this file", the harness reads it and sends the contents in, the model writes a patch, and the harness applies it. The same model in a different harness can feel much better or much worse.
Where the analogy breaks
-
The model does not run anything. It only asks, in text, and the harness decides whether to do it.
-
Two products with the same model can behave very differently. The harness chooses what the model sees and which tools it gets.
-
Much of the memory, safety checking and cost control lives here. When an agent forgets a rule, check what the harness put in the context, not just the model.