10  Summary

Context engineering is the discipline of managing the scarcest resource in AI systems: the context window. This book has established that:

1. Tokens are the currency — everything has a cost measured in tokens (Chapter 1)
2. Context windows are smaller than advertised — design for the smart zone, not the marketing number (Chapter 2)
3. The messages array has a fixed-cost structure — system prompt, harness, project context, and tools consume tokens before your conversation even begins (Chapter 3)
4. Tool calls are memory allocations — they grow context permanently within a session (Chapter 4)
5. Sub-agents are context isolators — delegate reads to sub-agents, keep writes centralized (Chapter 5)
6. Message passing is the architecture — design explicit protocols between context windows (Chapter 6)
7. Fresh context beats stale context — the Ralph Wiggum Loop (Requirements → Planning → Building) uses the filesystem as memory and the context window as disposable compute (Chapter 7)
8. Prevention beats compaction — avoid filling context rather than trying to compress it after the fact (Chapter 8)

The central insight: treat the context window as you would treat memory in a systems programming language — as a finite, precious resource that requires deliberate management.