HOW LLMs WORK FOR THE NON-SPECIALIST

WHY MODELS ARE STATELESS Stateless means that the model keeps no memory of earlier runs. Each time you call a model, it begins fresh, with only the text you send. If you tell it your name in one call and ask for your name in the next call without including the first message, it will not know. This surprises many beginners, because chat applications seem to remember the conversation. The secret is that the application resends the whole conversation each time. When you type your tenth message, the chat program packages your first nine messages and the model's nine replies together with the new message, and sends all of that as the input. The model is not remembering; it is rereading. Statelessness has big consequences for harness engineering. First, anything the agent must remember has to be stored by the harness and placed back into the context window when needed. Second, long projects eventually overflow the window, so the harness must decide what to summarize or drop. Third, two separate runs do not share what they learned unless the harness saves and reloads that information. This is why the harness is described as the model's memory, hands, and nervous system. In Part Five and Part Eleven you will study the techniques in detail.