langchain-middleware

Add human-in-the-loop approval to LangChain agents by configuring HumanInTheLoopMiddleware to pause execution before risky tool calls, requiring manual review before proceeding.

Per-tool interrupt policies allow you to specify which tools require human approval before execution. This ensures deterministic and auditable workflows by setting distinct approval rules for different actions.

Yes, a checkpointer like MemorySaver is required to pause LangChain agents for human approval. It maintains thread_id tracking to save the agent state while waiting for manual decisions.

Create custom middleware hooks for LangChain agents using configurable lifecycle events like before_model, after_model, wrap_tool_call, before_agent, and after_agent to inject custom logic.

Resume a LangChain agent after a human-in-the-loop interrupt by using command resume patterns, which route the human's decision back into the agent to continue execution from the paused state.

Use human-in-the-loop middleware for LangChain when agents perform cautious executions like sending emails, reading sensitive data, or modifying external systems where unintended actions are unsafe or costly.