Human-in-the-Loop

High-stakes actions — refunds, database writes, external emails — need human judgment. HITL embeds governance into the graph instead of bolting on a separate approval UI.

HITL is a co-pilot switch: the plane flies itself, but the captain must confirm before landing or changing course.

Scenario

Graph pauses before `send_email` node. Support lead edits the draft in a UI, clicks Approve; graph resumes and sends the edited version.

Solution

In Agent Frameworks, apply Human-in-the-Loop to this scenario: Graph pauses before `send_email` node. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Human-in-the-Loop (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Human-in-the-Loop happens in the code.

1app = graph.compile(checkpointer=saver, interrupt_before=["send_email"])2
3# First run — stops before send_email4app.invoke(input_state, config)5
6# Human edits state["draft"] in UI, then:7app.invoke(None, config)  # resumes and executes send_email

• Treating Human-in-the-Loop as a black box without evaluation
• Ignoring cost and latency in production
• Skipping error handling for langgraph human in the loop

Quick recap — the most important points from this module.