Crabhaus

What it is

How an agent turns one big goal into executable steps. Two poles: plan-then-execute, where it drafts a full step list before acting, and interleaved planning (the ReAct pattern), where it picks the next step fresh after each tool result.

How it works

Plan-then-execute wires a planner that writes an ordered task list, an executor that works through it, and a re-planner that revises when results diverge. Interleaved agents keep the plan one step ahead at all times, re-reading the latest observation each turn. Production harnesses blend both: a rough plan held in context as a todo list, revised whenever reality disagrees.

A concrete example

Ask a coding agent for a multi-file refactor and it writes a todo list first, then works item by item, re-ordering when tests fail. Deep-research agents decompose the question into sub-searches, fan them out in parallel, then synthesize.

Why it matters

Decomposition drives cost, latency, and reliability. Upfront plans let cheaper models run the easy steps, allow parallel fan-out, and give humans a checkpoint before the agent acts. Interleaved planning burns more frontier-model turns but recovers far better when the environment surprises.