Crabhaus

What it is

A sub-agent is a separate model instance an AI agent spawns to handle one delimited task in its own fresh context window. Isolation means the worker sees only the brief it is handed - not the parent's transcript - and the parent gets back only the result, not the worker's working.

How it works

The parent writes a short brief per sub-task. Each worker starts clean with that brief plus tools, does the heavy reading and searching on its own token budget, and returns a compressed summary; everything else it read is discarded. The parent's transcript grows by brief plus summary - small, cheap to re-send, clear - and isolated workers can safely run in parallel.

A concrete example

Asked why a test fails in a big repo, an agent spawns a search worker. The worker reads 60 files - roughly 80,000 tokens - and returns a 150-token answer naming the three relevant functions. Every later turn is priced and reasoned against those 150 tokens, not the dump.

Why it matters

Long transcripts hurt twice: models attend less reliably to facts buried mid-context ('context rot'), and every prior token is re-processed on every call, taxing each turn's cost and latency. Isolation is token hygiene - spend big context once in a disposable worker, keep the orchestrator lean.