Crabhaus

What it is

Two ways to wire an LLM system. A workflow runs the model inside developer-written control flow - code decides the next step. An agent flips it: the model picks its own next action, in a loop, until it judges the goal met.

How it works

Workflows chain fixed stages - classify, retrieve, draft, verify - so branching and retries are ordinary, testable code with near-fixed cost per run. An agent sees the goal plus its tool results so far, chooses the next call, observes, and repeats. Every self-directed step adds tokens, latency, and a fresh chance to go wrong, and errors compound across a long run.

A concrete example

A ticket pipeline that always classifies, pulls the account, drafts, then policy-checks is a workflow. A coding agent told to make a failing test pass - reading files, editing, re-running the suite until green - is an agent; nobody scripted that sequence.

Why it matters

Most agent reliability pain is unnecessary autonomy. If you can enumerate the steps, a workflow is cheaper, faster, and debuggable; agents earn their overhead only where the path can't be scripted. Anthropic's and OpenAI's builder guides agree: simplest solution first.