langchain-middleware

Human-in-the-loop approval for LangChain tool calls is added by applying HumanInTheLoopMiddleware to pause execution before dangerous actions. This requires a checkpointer-backed state with a thread_id to save the context and resume later.

You resume a LangChain agent after human approval by passing a Command object back to the graph. This Command pattern continues execution from the paused state, allowing the workflow to proceed with approved, edited, or rejected tool actions.

Yes, you can configure per-tool HITL policies in LangChain. This allows you to apply specific interception rules to individual tools, ensuring robust and auditable workflows where only dangerous or critical actions require human review.

Available middleware hooks to intercept LangChain agent actions include wrap_tool_call, before_model, after_model, before_agent, and after_agent. These hooks enable state inspection and custom validation logic before or after key execution phases.

Yes, a checkpointer is required to implement human-in-the-loop workflows in LangChain. It provides the state persistence needed to pause execution, maintain a thread_id for context, and reliably resume the agent after human decisions.

LangChain agents execute tool calls without pausing when HumanInTheLoopMiddleware is not applied. Without per-tool interrupt configurations and a checkpointer-backed thread_id, the workflow bypasses human oversight and runs actions automatically.