Crabhaus

What it is

A deliberate pause built into an agent workflow: the agent stops before executing a specific action and waits for a person to approve, edit, or reject it. It's a runtime control on what an agent may do - separate from anything that happens during training.

How it works

The agent proposes an action, the framework checkpoints state, and execution pauses until a human answers - in LangGraph, interrupt() saves graph state and a resume Command feeds the decision back. Placement is the design question: reversible actions (reading, drafting, branch edits) run freely; irreversible or outward-facing ones (sending, spending, deleting, deploying) wait for approval.

A concrete example

A team coding agent iterates on code in a shared workspace all day, but the deploy only goes out after a human signs off - the pattern Slack shipped this week in Slack Code. In a personal agent: it drafts replies freely, but sending an external email needs your click.

Why it matters

Agents fail in ways tests don't catch, and one bad irreversible call can outweigh a thousand good ones. A well-placed gate turns a catastrophic error into a recoverable one - which is exactly what lets you grant broad autonomy everywhere else, plus an audit trail of who approved what.