WHAT A HARNESS IS

A FORMAL DEFINITION An agent harness is the software infrastructure that surrounds a language model so that the model can act as a reliable agent. Put in a single sentence: the harness is everything in an AI system except the model itself. Let us unpack that sentence. The model, as you learned, is a trained network of numbers that turns an input of tokens into an output of tokens. It cannot open a file, send an email, or run a program. It cannot count how much money it has spent. It cannot remember yesterday. It can only produce text. The harness is the collection of ordinary software, written in an ordinary programming language, that supplies everything else. It sends the model its input. It reads the model's output and decides what the output means. If the output is a request to use a tool, such as "read the file named report.txt," the harness checks whether that is permitted, actually reads the file, and gives the contents back to the model on the next turn. If the output is a final answer, the harness delivers it to the user. Along the way the harness records what happened, keeps track of spending, handles errors, and stops the run when the work is done or when a limit is reached. The word harness comes from the idea of a horse harness, the straps and fittings that let a powerful animal pull a cart in a useful direction. A strong horse without a harness is impressive but not useful. With a harness, the same strength is directed, controlled, and connected to the cart. An AI model is similar: powerful, but only useful for sustained work once it has been harnessed.