Context engineering

Context engineering is the quality control discipline for what an agent is allowed to “know” at each step. Agents have no inherent memory; the factory must supply an explicit, persistent context layer—often Markdown, JSON, or YAML on disk, plus search (rg) and indexes.

Good context engineering balances completeness (enough truth to implement) with focus (scoped loading so tokens are not wasted on irrelevant files). When hierarchies grow messy, teams reorganize knowledge for retrieval—sometimes called genrefying, like reclassifying a library.

In broader industry usage, the same phrase also names the strategic design of what goes into (or stays out of) the context window—including compression, retrieval, and memory tiers. That view aligns with RAG, system prompts, and context windows as first-class constraints.

See also: Seed manufacturing system, Validation harness, Context window, RAG (retrieval-augmented generation).