Crabhaus

What it is

Context engineering is deciding what goes into a model's context window each turn - system instructions, retrieved documents, tool definitions, history, memory - under a fixed token budget. It reframes 'prompt engineering' (wording one prompt) as a systems problem: for an agent running many steps, the real question is which information to assemble and what to leave out.

How it works

A context window is finite, and models degrade as it fills - attention thins and middle content is underweighted, the 'lost in the middle' effect. So tokens are a budget: a compact system prompt, small tool schemas, snippets retrieved just-in-time rather than dumped wholesale, distilled memory, and the transcript. Long runs compact stale outputs so the window keeps only the highest-signal set.

A concrete example

A coding agent fixing a failing test doesn't paste the whole repo. It pulls the failing test and the source it exercises, the error traceback, and its edit and run-tests tool schemas, and summarizes what it already tried. After ten steps, earlier raw outputs are compacted into a short 'what I've learned' note so the window stays focused instead of hauling every command's output forward.

Why it matters

For agent builders, retrieval and context assembly usually decide quality and cost more than the base model. A bloated context is slower, pricier, and often less accurate than a lean one, and a cache-friendly layout cuts inference cost. As agents run longer, 'just enlarge the window' stops working; curating what the model sees each step keeps it reliable.