WTF is Context Engineering?
Deciding exactly what goes into a model's context window at each step: instructions, tools, data and history.
TL;DR
Prompt engineering is about wording one prompt well. Context engineering is the bigger job of curating everything the model sees on every turn of a long agent run: which documents, which tool results, what to summarise, what to drop. Most agent failures are context failures, not model failures.
How it works
Techniques include just-in-time retrieval (load data when needed instead of up front), compaction (summarising old turns when nearing the context limit), structured note-taking or external memory files, sub-agents that explore in their own context and return condensed results, and careful tool design so results are compact. The core constraint is that context is a finite attention budget: more tokens can make performance worse ("context rot").
Origin
- Coined / introduced by
- Popularised by practitioners
- When
- Jun 2025
- Story
- Gained wide use in mid-2025 after posts by Shopify's Tobi LĂĽtke and Andrej Karpathy; AI labs then published guides using the term.