Vol. VI · Context & Memory · VI.1
Context Engineering
The discipline of deciding, every turn, what tokens the model sees. More leveraged than prompt wording, more neglected than tool choice.
8 min read
humanlayer/12-factor-agentsmuratcankoylan/Agent-Skills-for-Context-Engineeringaddyosmani/agent-skillsAider-AI/aiderletta-ai/letta
Doctrine
- Measure the window: token counts per section, every turn, in traces.
- Prefetch what you will need (appendix factor 13) rather than hoping the model asks.
- Design eviction: what is pinned, what is summarized, what is filed, what is dropped.
A field, not a slogan
muratcankoylan's Agent-Skills-for-Context-Engineering and Addy's skills treat this as checklists: isolate multi-agent contexts, cap tool output, pin the goal, file the rest. Aider's repo map is context engineering for codebases. Letta is context engineering for lifelong agents.
If you only steal one habit: log the constructed prompt, not just the user message. You cannot debug a window you never saved.
Anti-patterns
- RAG everything. Retrieval is one tactic, not the discipline.
- Ignoring that tool outputs are the usual window killer.