How context engineering actually works.

It's the discipline of managing what inputs actually reach an AI model, beyond just the prompt. The post describes assembling retrieval results, tool outputs, and memory into the context window, and, crucially, deciding what gets discarded so the most relevant information is what the model ends up processing.

Retrieval (which pulls documents), tools (which return live results), and memory (which carries prior turns and durable facts about the user). The post calls these the easy part because they map onto systems engineers already understand, a search index, an API, and a database.

Because the window has a fixed size and every token from every source competes for the same finite room. When the three sources together produce more than fits, and they almost always do, something must be cut. Deciding what to cut isn't a side effect; it's the central act of context engineering.

Compaction is the policy that decides what gets summarized, demoted, or dropped as the window fills, drawn as the dashed orange loop in the diagram. The post argues teams wire retrieval, tools, memory, and the model straight through and wonder why quality degrades, because they never designed the discard.

A retrieval bug shows up as a missing fact, which is relatively easy to spot. A compaction bug appears as a model that was right an hour ago and is confidently wrong now, much harder to catch because nothing obviously failed; the system simply discarded something it should have kept.

As code, with its own evals, not as an afterthought tuned by feel. The post applies its "boring wins" principle here: a predictable, well-tested compaction step beats a clever one nobody can reason about at 2am. Designing what to leave out is treated as a first-class engineering task.

Because teams forget to design the discard. The post says they draw retrieval, tools, memory, and the model, wire them straight through, and never define compaction, the policy for what gets summarized, demoted, or dropped as the window fills. Without it, the model silently loses things it needed.

It means context engineering is mostly the engineering of what to leave out. Before the model processes anything, the assembly stage has already chosen which retrieval results, tool outputs, and memories survive the budget. The model never sees the discarded material, so that decision largely determines output quality.