WHAT A HARNESS IS

THE MODEL AS A STATELESS REASONING ENGINE, THE HARNESS AS THE RUNTIME ENVIRONMENT Two phrases deserve careful attention. The first is stateless reasoning engine. Stateless, you will recall from Chapter 2, means the model keeps no memory between calls. Reasoning engine means that what the model does well is to think about the text in front of it and produce a sensible continuation, which may include a plan, a decision, a piece of code, or a request to use a tool. The second phrase is runtime environment. A runtime environment is the surroundings in which a program actually runs, including the resources it can reach and the rules it must obey. For example, a web browser is the runtime environment for a web page, because it provides the screen, the network connection, and the safety rules. The harness plays that role for the model. It decides what the model can see, what it can do, how long it can work, and what happens when it makes a mistake. This division of labor is powerful because each side does what it is best at. The model is flexible and good with ambiguity, meaning unclear or open-ended situations. The harness is exact and good with rules, meaning situations where the same input must always produce the same behavior. When you assign a task to the wrong side, trouble follows. If you ask the model to remember a list of 500 completed items, it will eventually forget or invent entries, while a harness can store the list perfectly in a file. If you try to write rigid code that interprets every possible phrasing of a customer complaint, you will fail, while a model reads the complaint easily. A central skill of harness engineering is putting each responsibility on the correct side of the line. A helpful rule of thumb is this. Anything that must be exact, remembered, limited, secured, or recorded belongs to the harness. Anything that requires interpretation, judgment among fuzzy options, or the generation of natural language belongs to the model.