THE THREE DISCIPLINES AROUND THE MODEL

CONTEXT ENGINEERING Context engineering is the practice of curating all of the information placed in the context window across the turns of an interaction. Recall from Chapter 2 that the context window is the model's desk, and that anything off the desk is invisible. Context engineering is the craft of deciding what goes on the desk at each step. That includes the standing instructions, the recent conversation, selected documents, the results of tool calls, retrieved facts from a knowledge store, and summaries of older material. It also includes deciding what to leave off. An overloaded desk is a problem for several reasons. It costs more, it slows the response, and it can distract the model with irrelevant details. Imagine a lawyer preparing for a meeting. A good assistant does not drop the whole filing cabinet on the lawyer's desk. The assistant chooses the five documents that matter for this meeting, puts a short summary on top, and keeps the rest within reach. That is context engineering. In an agent that works for hours, context engineering becomes demanding, because the history grows until it no longer fits. The harness must keep, trim, summarize, and fetch information so the model always has what it needs. Context engineering is so central that it forms one of the seven layers of the ETCLOVG framework, covered in Part Five.