Shrink and curate context

Since context is paid for on every loop iteration, controlling it is controlling cost:

• Reference, don't embed (Chapter 5). Don't paste documents into the prompt; let the model fetch via tools, paying for content only when used.
• Distill tool results. Cap long payloads, return decision-relevant fields, offer a "get more" tool. A giant unfiltered tool result is re-sent on every subsequent iteration, paying for it repeatedly.
• Curate history. Use a recent window or summarization rather than re-sending an ever-growing transcript.
• Keep prompts focused. Load task-specific instructions on demand instead of carrying every possible instruction in the system prompt on every call.