## 1. The Working Premise ### Context is the agent's attention surface, not its memory ### Most agent failures are context-engineering failures, not model failures ## 2. Context Curation ### Just-in-time retrieval beats just-in-case loading ### Compress on the way in (titles → summaries → full text) ### Drop stale tool outputs before they pollute the window ## 3. Memory vs Context ### Memory persists across sessions ### Context is per-turn ### Re-loading from memory is the bridge ## 4. Long-context Failure Modes ### Lost in the middle ### Conventions drowned by intermediate outputs ### Phantom progress after compaction ## 5. Anti-patterns ### Kitchen-sink system prompts ### Re-summarising the summary ### Loading everything "in case" ## 6. Operational Practices ### Externalise durable decisions to files ### Restart sooner than instinct suggests ### Use sub-agents to keep verbose tool output out of the main thread ### References - https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents