Context Engineering
The practice of designing and managing everything that goes into a model's context window — instructions, retrieved documents, tool outputs, and memory — so the model has exactly the information it needs, and nothing it doesn't.
Context engineering is the discipline of assembling a model’s input. Where prompt engineering focuses on how a request is phrased, context engineering treats the entire context window as a budget: system instructions, conversation history, retrieved documents, tool schemas, tool results, and memory all compete for tokens.
The shift matters because modern AI systems — especially agents — generate most of their own context. A coding agent that reads files, runs tests, and retries builds a context that no human hand-wrote. If that context fills with stale tool output or irrelevant retrieval, the model’s reasoning degrades even when the prompt itself is perfect.
Common techniques include:
- Compaction: summarising older turns so recent, relevant information fits.
- Retrieval discipline: injecting only documents that pass a relevance threshold.
- Structured state: keeping durable facts in files or a database instead of replaying them in chat.
- Sub-agents: delegating side quests so the main context stays clean.
The term gained traction alongside agentic AI, where context overflow and goal drift are among the most common causes of failure on long-horizon tasks.