Crabhaus

What it is

Treating an agent's context window as a finite budget split deliberately across competing claims: system instructions, conversation history, retrieved documents, tool results, working notes, and space reserved for output. Instead of letting whatever arrived last fill the window, the agent decides what earns a place on every call - the way a program budgets memory.

How it works

Every call re-reads the whole context, so the budget is spent per step. Systems combine pinning (instructions always stay), compaction (past a watermark, old turns are summarized and dropped), pruning (bulky tool results truncated once used), just-in-time retrieval (keep IDs, fetch content when needed), external notes files, and a reserved output allowance.

A concrete example

A coding agent on a long refactor pins its plan, lets a 200-line test log in once, then swaps it for 'tests: 3 failures in auth module (see notes file)'. At 80% full it compacts 60 earlier turns into a half-page brief of decisions and files touched, keeping recent turns verbatim - so the newest compiler error is never crowded out.

Why it matters

Long-running agents rarely die at the hard window limit; they degrade earlier, as attention spreads across stale logs and re-reading an overstuffed context every call compounds cost and latency. A budget keeps the highest-value evidence - current goal, recent results, binding constraints - in the working set, which is what keeps behavior coherent over hundreds of tool calls.