How do I test a LangGraph agent that pauses for human approval?

Resume the paused graph with Command(resume=...) on the same thread_id the run paused under, and hand it the approve, edit, reject, or respond value your test wants instead of waiting for a person to click anything. LangGraph’s human-in-the-loop pattern treats that value the same way it treats any other resumed state: the graph picks up exactly where interrupt() left it, with your supplied decision standing in for whatever a person would have given it.

That turns every branch of an approval flow into something a test can drive directly: script an approve to check the happy path, a reject with a message to check how the agent handles pushback, or an edit to check whether it actually uses the changed value instead of the one it first proposed. It’s the same problem a CI gate has to solve for any check that would otherwise need a person watching, aimed at the one node built to wait for one.