Crabhaus

What it is

An agentic setup where a model runs a long investigation on its own: it breaks a question into sub-queries, searches the web, reads what it finds, and loops - search, read, re-plan - until it can write a cited report. OpenAI's Deep Research and Claude's Research feature are the best-known examples; open-source clones followed within weeks.

How it works

A lead agent plans breadth versus depth: fan out across independent sub-questions, then drill into promising branches. Each search-read cycle feeds back into the plan, so the run is path-dependent, not a fixed pipeline. Anthropic spawns parallel subagents with separate contexts; OpenAI uses one browsing-optimized reasoning model. Claims tie back to URLs visited; runs burn ~15x a chat's tokens.

A concrete example

Ask one for every EU AI Act obligation hitting general-purpose model providers in 2027 and how vendors are responding. It splits the job into regulation text, timeline, and vendor moves, searches each branch, reads primary documents, discards weak pages, and returns a sectioned report with per-claim links - breadth-first work where parallel subagents beat one sequential searcher.

Why it matters

This is becoming the default interface for 'go find out X' at work, and it changes the cost model: a research run spends a large token budget, so it pays off only when the answer is worth more than the compute. For builders, the planner-plus-parallel-searchers-plus-citation-tracking shape is the current reference design for long-horizon information tasks.