Context engineering

Ask Claude about one official copy and it usually answers well. Give the same agent the whole contract pack (official copies for three titles, an 80-page lease, a local search, a drainage and water search, an environmental report, the TA6 and TA7, the management pack) and ask the same question, and the answer often gets worse. It might miss the restriction in the proprietorship register, or attribute a covenant to the wrong title. The model hasn't become less capable. What changed is what it was looking at. Context engineering is the practice of deciding what an agent sees, when, and in what form. Anthropic describes it as "curating and maintaining the optimal set of tokens (information) during LLM inference" (Rajasekaran et al., 2025).