Crabhaus

What it is

The recurring shapes for coordinating multiple LLM agents: orchestrator-worker (a lead agent splits the task across parallel sub-agents), handoffs (one agent passes the whole conversation to a specialist), and debate (independent agents answer, then critique or vote).

How it works

An orchestrator plans, spawns workers with fresh context windows, and synthesizes their short summaries. Handoffs transfer control, not work - one agent active at a time, each with a small, focused prompt. Debate buys error-checking by running one question through independent contexts and reconciling disagreements with a judge or a vote.

A concrete example

Anthropic's research feature: a lead agent plans, parallel search sub-agents explore, and the lead synthesizes. Anthropic reports 90.2% better results than a single-agent Claude Opus 4 baseline on its internal research eval - at roughly 15x the tokens of a normal chat.

Why it matters

Coordination buys context isolation, wall-clock parallelism, and independent perspectives; it costs tokens, latency, and new failure modes like duplicated work. Orchestration pays on tasks that decompose cleanly - research, review, migration sweeps - and mostly multiplies the bill on tightly coupled work.