Context on a budget: what to keep, compress and cut

Every request re-sends everything the model needs: instructions, tool definitions, evidence and the conversation so far. That is your context, and it is a budget. Spend too little and the model lacks the facts; spend too much and you pay for tokens that make answers worse, slower and more expensive. Context engineering is deciding, every turn, what goes in, in what order, and what comes out. You allocate by priority, prune what has gone stale (old tool output, duplicate chunks), compact long history into a summary that keeps IDs, numbers and commitments verbatim, fetch detail just in time instead of stuffing it up front, and put everything static first so the provider can cache it. Then you measure it per turn, because context bloat is how agents that were fine in week one become slow, expensive and wrong in month three.

• The window is a ceiling, not a target. Usable budget is the window minus output reserve and safety margin.
• Agent loops make input cost grow roughly quadratically with turns, because history is re-sent.
• Pruning is cheap and safe; compaction is lossy. Prune first, compact second, and pin what must never be summarised.
• Static first, dynamic last. Stable prefixes are what prompt caching rewards.
• If you do not log tokens per category per turn, you cannot explain a regression.